What changed in robot arm hardware, and when

This document records the hardware events of roughly the last two years: which arms, hands, sensors and computers appeared, which ones changed price, which companies were bought, and which announcements have still produced nothing you can buy. It is the document to read before spending money.

It is deliberately an events document. It does not try to explain why hardware moves in the direction it moves, because what is changing in robot manipulation already does that, and the two overlap on purpose. If you want the mechanism, read that one. If you want to know what happened and on what date, read this one.

It is also not a catalogue. Grippers and the hardware around them already gives the payloads, strokes, forces and drivers for every gripper family, and this document links to it rather than repeating it. Where the two documents touch, that one tells you what a product does and this one tells you when it arrived and what it replaced.

Who this is for#

Someone with a budget and a decision to make. A person deciding whether to buy a cheap arm or a capable one, a lab deciding whether last year's quote is still good, or anyone who has read an announcement and wants to know whether the thing in it exists.

Every figure below was read from the manufacturer's own page, its own online store, or its own repository, fetched in September 2026, and the link is given so you can check it. Historical prices come from dated snapshots in the Internet Archive or from dated commits in a public repository, and those links are given too. Where a vendor publishes no price, the document says so instead of quoting a reseller. Where a page could not be read at all, the document says that as well.

Contents#

1. How to read this document
2. Collaborative and industrial arms
3. Research and low-cost arms
4. Humanoid arms and hands
5. Grippers and multi-finger hands
6. Tactile and force sensing
7. On-robot compute
8. The category that did not exist: teleoperation hardware
9. What did not change
10. How to check an announcement yourself

1. How to read this document#

1.1 The four maturity labels#

Robot hardware announcements are unusually unreliable. A press release, a video and a product you can order are three different things, and the industry uses the same vocabulary for all three. Every item in this document therefore carries one of four labels, and the label is a statement about evidence rather than about quality.

Shipping means you can buy it today. Where the vendor publishes a lead time, the lead time is given, because a product with a twelve-week lead time and a product you can have next week are different purchases.

Announced with a date means the vendor has committed publicly to a date. An order page that takes a deposit counts here, not as shipping, because a deposit is a promise rather than a delivery.

Demonstrated means somebody has shown it working, usually in a video, and there is no way for you to obtain one. Much of the humanoid industry sits here.

Vapour means it was announced, no date was given, and nothing has been shown. This label is not an insult. It is a description of the evidence, and several respectable companies have products in this state.

1.2 Why prices are so hard to quote#

Most of the robot arm industry does not publish prices, and the split is not random. It follows the sales channel.

A vendor that sells through system integrators publishes no price, because the price depends on the integration. A vendor that sells to individuals runs a web shop and publishes everything. That single distinction explains almost every gap in this document.

Of the collaborative arm vendors, none of Universal Robots, Franka Robotics, Kinova, Doosan Robotics, Techman, Elite Robots, JAKA or Dobot publishes a price on its own website. Every one of those sites was fetched in September 2026 and every one routes you to a quote form. The exceptions are UFACTORY, Unitree, Trossen Robotics and AgileX, all of which run a shop with prices on it, and all of which appear repeatedly below for exactly that reason.

This produces a bias you should be aware of while reading. The numbers in this document are concentrated in the parts of the market that sell to individuals, because that is where numbers exist. It is entirely possible that industrial arm prices moved and nobody can see it. The honest position is that the published evidence covers the cheap end well and the industrial end barely at all.

One more warning about prices from low-cost vendors. A repeatability figure of 0.1 mm on a two-thousand-dollar arm is the vendor's claim, measured by the vendor, under conditions the vendor does not state. The same figure on a Universal Robots datasheet is backed by a test standard. Treat the numbers below as what the vendor published, which is what they are, and not as independently measured.

2. Collaborative and industrial arms#

A collaborative arm, usually shortened to cobot, is an arm designed to work beside people without a safety cage, by limiting its own speed and force. This is the category that Universal Robots created and that everybody else entered.

2.1 What a mid-range arm cost in 2024, and what it costs now#

This is the headline question, so it goes first, and the answer is not the one the folklore gives.

Before, in mid-2024, the cheapest capable six-axis arm with a published price was the UFACTORY xArm 6: 5 kg payload, 700 mm reach, 0.1 mm repeatability. The Internet Archive's snapshot of UFACTORY's own product page taken on 24 June 2024 gives the price as US$8,399.00 to US$8,594.00.

What changed is that nothing changed. The same page fetched on 23 September 2026 gives exactly the same price, US$8,399.00 to US$8,594.00, for exactly the same specification. A snapshot from January 2025 gives the same figure again. Three readings spanning twenty-seven months, and the number did not move once.

How that came about is the ordinary economics of a mature product. The xArm 6 is a harmonic-drive arm built in volume against a stable bill of materials, and there was no competitive event in that band that forced a cut. The Chinese entrants that did arrive attacked a different band, which is section 3.

Why it matters is that "arms got cheaper" is folklore when applied to this part of the market. If you were quoted a mid-range industrial arm in 2024 and waited for the price to fall, you waited for nothing.

What this comparison still cannot tell you is what happened to the arms nobody publishes a price for, which is most of the industrial market. A UR10e or a Franka Research 3 may well have moved. There is no published evidence either way, and a reseller's figure is not evidence.

The full set of UFACTORY prices, read from the shop's own product feed on 23 September 2026, is worth having because it is one of the few complete published price lists in the industry. Read it as: the model, what it lifts, and what UFACTORY charges for it.

Two entries in the same feed are not arms and are more surprising than the arms are. UFACTORY publishes $3,000 for a six-axis force-torque sensor and $1,999 for its xArm Gripper G2. Section 6 comes back to the first of those, because a published three-thousand-dollar force-torque sensor is a genuine change.

Status: Shipping.

2.2 Universal Robots announced a new platform on 14 September 2026#

Before, Universal Robots sold the e-Series, launched in 2018, extended upwards with the UR20 and UR30 and then with the UR15. The UR18 was added on 6 October 2025, according to Universal Robots' own news centre. Every one of those is the same architecture with a different arm on the end.

What changed is that on 14 September 2026 — nine days before this document was written — the same news centre announced "Universal Robots unveils Gen 7, a new platform for industrial automation and physical AI deployment", described as "out-of-the-box AI-ready". The arms in it are branded the g-Series, and Universal Robots' g-Series page calls it the company's "7th generation robot family within the Gen 7 platform" and lists three models: the UR10g-1750 at 8 kg payload and 1750 mm reach, the UR17g-1300 at 15 kg and 1300 mm, and the UR18g-950 at 18 kg and 950 mm.

The UR10g-1750's own product page gives the figures that matter: an 8 kg payload, 10 kg in an extended configuration, a pose repeatability of ±0.08 mm to ISO 9283, a maximum tool speed of 5 m/s and a robot mass of 44.7 kg. The launch also brings a new CB7 controller, which Universal Robots describes as carrying 40 per cent more compute in a 30 per cent smaller footprint, a TP7 teach pendant more than 20 per cent lighter, and an SP7 flange panel.

How it was achieved is stated plainly on that page, and the interesting part is not the arms. The g-Series brings "power, data, and safety connectivity directly to the tool flange", with integrated tool flange inputs and outputs at 24 or 48 volts, 5 A peak and 3 A continuous, and 1 gigabit Ethernet "for cameras and sensors". Universal Robots has moved the wiring for a camera and a smart gripper inside the arm. That is a response to the fact that almost every modern cell bolts a camera to the wrist and then runs a cable down the outside of the robot.

One thing Gen 7 explicitly did not bring is worth recording because it is widely assumed. The g-Series does not have joint torque sensors. The specification page lists a force-torque sensor at the tool flange and gives no per-joint torque figure at all. A separate 2025 Universal Robots feature called Direct Torque Control, which gives "real-time control of all the joints in a UR cobot at 500 Hz", is a torque command interface and says nothing about how torque is measured. Section 9 returns to this.

Why it matters is that the cable you no longer have to route is one of the larger hidden costs of integrating an arm, and one gigabit at the flange is enough for a wrist camera without a separate umbilical.

What it still cannot do is be priced. Universal Robots publishes no price for any of it. Nor could this document verify a shipping date: the news centre index page carries the headline and the date, but the individual press release could not be fetched, because the site serves its article pages from a client-side application that returns nothing to an automated request.

Status: Announced with a date, for the announcement itself on 14 September 2026. Whether the g-Series arms are shipping could not be confirmed from a page that resolves.

A second Universal Robots item belongs here because it is compute rather than mechanics. The UR AI Accelerator is a bundle of an "Embedded NVIDIA Jetson Orin AGX 64GB compute box" with an Orbbec Gemini 335Lg depth camera, built on PolyScope X with "full ROS2.0 support", and the page says it "is now available for order". No price. The reason it matters is covered in section 7: an arm vendor shipping a named Jetson module as a catalogue item is a change in what an arm is expected to come with.

There is also a software-side event with hardware consequences. The same news centre records, on 16 March 2026, "Universal Robots and Scale AI launch imitation learning system", unveiling the UR AI Trainer at GTC 2026 and describing data generated in training cells. A training cell is hardware. Section 8 returns to this.

2.3 Franka is now part of Agile Robots#

Before, Franka Emika sold the Panda and then the FR3, and was the default research arm for anybody who wanted torque sensing in every joint at a price a university could reach.

What changed is ownership. Franka Robotics' own site states that the company was "Founded in 2016, it is part of Agile Robots since 2023". The product page now lists the Franka Research 3 alongside the "Diana 7 by Agile Robots", so the two catalogues have merged. It also lists a set of prototypes: FR3 Duo, Mobile FR3 Duo, a Tactile Mobile Robot, and Franka GELLO and GELLO Duo.

How it happened is a matter of public record only in outline, and this document does not have a verified account of the insolvency and sale, so it does not give one. What the vendor's own page states is the sentence quoted above, and that is what is reported here.

Why it matters is practical. If you are reading a paper from 2022 or 2023 that used a Franka Panda, the company that made it no longer exists under that name, and the arm in the paper is two generations back. If you are buying, you are buying from Agile Robots.

What it still cannot do is tell you a price. Franka publishes none, and the "Buy now" control is a quote request.

The GELLO prototypes deserve a note of their own. GELLO is an open-source leader arm for teleoperating a robot arm by hand, and a major arm vendor now lists its own version as a product. That is section 8.

Status: Franka Research 3, Shipping. FR3 Duo, Mobile FR3 Duo, Tactile Mobile Robot, Franka GELLO and GELLO Duo are labelled prototypes by the vendor, which places them at Demonstrated.

2.4 The Chinese entrants, and what can actually be checked#

The claim you will meet everywhere is that Chinese manufacturers reset the price floor for collaborative arms. This document can confirm part of that claim and not the rest, and the distinction matters.

What can be confirmed is that Chinese vendors now dominate the bands below $10,000 with published prices. UFACTORY, Unitree, AgileX, Dobot and Elephant Robotics all publish, and their arms are all Chinese-built. Section 3 has the numbers, and they are genuinely low.

What cannot be confirmed is that they pushed down the price of a UR10e or a Franka. JAKA, Dobot, Elite Robots and Techman were all fetched in September 2026 and none of them publishes a price for anything. There is no published series a reader can inspect, so a statement that Chinese competition halved industrial cobot prices would be folklore repeated with confidence. It may be true. It is not checkable from vendor sources.

There is, however, one piece of indirect evidence that the pressure is real, and it is a court docket rather than a price list. Teradyne, which owns Universal Robots, brought a copyright action against Elite Robots Deutschland in February 2026, and the Regional Court of Hamburg issued a preliminary injunction in Teradyne's favour on 21 April 2026, barring Elite Robots Germany from offering the software at issue and requiring it to disclose which customers had been supplied. In August 2026 Teradyne sued JAKA Robotics GmbH at the Unified Patent Court in Copenhagen, case UPC-CFI-0003057/2026, over three Universal Robots patents covering touchscreen robot programming, joint safety brakes and joint construction. Neither case has reached a final ruling.

How to read that is a matter of judgement rather than fact, and this document will state its judgement plainly and mark it as such. An incumbent that competes on price cuts prices. An incumbent that goes to court is defending a position that price alone is no longer holding. That is the clearest available signal that the Chinese entrants have become genuinely competitive, and it is a signal rather than a measurement.

Two other Teradyne numbers are worth recording because they contradict the common story that the cobot market is collapsing. Teradyne's own second-quarter 2026 results report $100 million in robotics revenue, its first hundred-million-dollar quarter, up 33 per cent year on year and the fifth consecutive quarter of growth. Robotics is nonetheless down to 8 per cent of Teradyne's total revenue from 12 per cent a year earlier, because the rest of the company grew faster. The robotics business is growing; it is simply growing inside a company whose semiconductor test business is growing faster.

One correction while this subject is open, because it is repeated constantly and it is wrong. Teradyne Robotics is Universal Robots, Mobile Industrial Robots and Energid. Robotiq is not part of it, and neither is OnRobot. No acquisition of Robotiq appears anywhere in the public record, and Robotiq's own about page says nothing about ownership either way.

3. Research and low-cost arms#

This is where the money actually moved, and it moved in a specific direction: not the cheapest arms getting cheaper, but the gap between $200 and $8,000 filling up with things that did not exist.

3.1 The hundred-dollar arm stopped getting cheaper#

Before, in early 2024, the cheapest credible learning arm was the Koch v1.1, a printed five-axis arm built around Dynamixel XL330 servos. Its repository puts the leader arm's bill of materials at $199 in the United States. The repository is Apache-2.0 and has had no commits since September 2024.

What changed first was the SO-100, which replaced Dynamixel servos with cheaper Feetech ones. The SO-ARM100 repository's README as it stood on 25 October 2024 gives $241 for a leader and follower pair in the United States and 244€ in Europe, and $127 for a single follower arm, 128€ in Europe.

What changed since is almost nothing. The same repository today gives $229.88 for the pair and €226.3 in Europe, and $121.94 for a single follower arm, €124.3 in Europe. That is a fall of about five per cent over twenty-three months. The SO-101 superseded the SO-100 — the README calls the SO-100 documentation deprecated — but the change was mechanical, chiefly the leader arm's gearing, not a price cut. As of September 2026 the SO-101 is still the current version in the official repository, and there is no SO-102 in it.

How the original drop was achieved is worth stating because it explains why it did not repeat. The Koch used Dynamixel XL330 servos at $24 each. The SO-100 used Feetech STS3215 servos, which WowRobo sells today at $15.99 each, or $13.99 each in packs of ten. Six servos is most of the cost of a printed arm, so switching servo supplier was the whole saving. There is no second saving of that kind available, because the remaining cost is the servos themselves.

Why it matters is that the hundred-dollar arm is a floor rather than a trend. If you are waiting for a fifty-dollar arm, the bill of materials says it is not coming from this direction.

What it still cannot do is a great deal. These arms have no joint torque sensing, no meaningful payload, and repeatability nobody publishes. They exist to collect demonstrations and run small policies, which is exactly what the learning path uses them for.

Status: Shipping. The repository licence is Apache-2.0, which is a software licence applied to a hardware project — see section 3.5 for why that distinction is worth checking every time.

3.2 The real change was the kit ecosystem#

Before, building an SO-100 in 2024 meant owning a 3D printer, sourcing servos from Alibaba, and assembling the arm yourself. The bill of materials was the product. That is a large barrier disguised as a low price, and it is the reason the hundred-dollar arm was less accessible than its headline suggested.

What changed is that resellers appeared. The SO-ARM100 repository now lists ten of them, in Switzerland, the United States, China, Japan, South Korea and the European Union, selling printed frame kits, electronics kits, complete kits and assembled arms. WowRobo publishes $199 for the do-it-yourself kit, $259 and $299 for assembled versions. Seeed Studio lists an SO-ARM100 servo motor kit at $200, $196 each in tens, without printed parts, and showed as out of stock when fetched.

How it happened is straightforward commerce. An open design with a large user base and no trademark restriction is an invitation to manufacture, and the Apache-2.0 licence permits exactly that. The design's authors get no royalty and the buyer gets an assembled arm.

Why it matters is that the effective cost of entry went up and the effort went down, and the second matters more. An assembled SO-101 at $299 against a $122 bill of materials is a $177 premium for not owning a printer and not spending a weekend. For most people that is the transaction that made the arm real.

What it still cannot do is guarantee anything. These are third-party manufacturers of somebody else's open design. There is no warranty position, no support commitment, and no guarantee that two sellers' arms are dimensionally identical, which matters when a trained policy is supposed to transfer between robots.

Status: Shipping, from multiple vendors, with stock varying.

3.3 The two-thousand-dollar arm filled the gap#

Before, the choice was a printed arm at a few hundred dollars with no payload, or an industrial arm at $8,000 and up. There was very little between them, and the gap was the single most awkward fact about equipping a manipulation lab.

What changed is that AgileX's PiPER sits in it at $1,999: six degrees of freedom, 1.5 kg payload, 626 mm reach, an arm weighing 4.2 kg, and a stated repeatability of 0.1 mm. AgileX's own store publishes the price and states Python, ROS 1 and ROS 2 support. A PiPER-X variant is listed at the same $1,999.

How it was achieved is the same answer as everywhere in this section: an integrated actuator module built in volume for the Chinese domestic market, sold through a Shopify storefront that skips the integrator channel entirely.

Why it matters is that 1.5 kg of payload with a real serial interface is enough to do actual manipulation work — it will hold a drill, a bottle or a tool — at a price a single researcher can approve. The band between a toy and an industrial robot stopped being empty.

What it still cannot do is prove its own numbers. The 0.1 mm repeatability figure is stated without a test standard or measurement condition, and this document found no independent measurement of it. Treat it as a claim.

Status: Shipping.

3.4 Trossen replaced the ALOHA kit, and it did get cheaper#

This is the clearest like-for-like price fall in the document, because one vendor published both numbers.

Before, the standard bimanual imitation-learning rig was Trossen's ALOHA Stationary kit: two ViperX 300 S arms as followers and two WidowX 250 S arms as leaders. The Internet Archive's snapshot of 2 April 2024 prices it at $29,999.95, and a snapshot from March 2025 gives the same figure. The follower arm was the expensive part: a snapshot of the ViperX 300 page from 18 February 2025 gives $6,129.95 for 6 degrees of freedom, 750 mm reach and a 750 g payload.

What changed is that Trossen retired the line and replaced it. The ALOHA kits page now says "ALOHA IS NOW TROSSEN AI". Trossen's Stationary AI is $23,995.95 and contains "2X Leader-Follower WidowX AI Arm Pairs", with an average production time of two to three weeks. The WidowX AI arm sells on its own at $4,545.95 for the base version, $4,685.95 as a leader and $4,995.95 as a follower, with 6 degrees of freedom, 1.5 kg payload, 700 mm reach and 1 mm repeatability.

So the kit fell from $29,999.95 to $23,995.95, a drop of twenty per cent, and the per-arm payload doubled from 750 g to 1.5 kg at the same time. The single arm fell from $6,129.95 to $4,545.95, a drop of twenty-six per cent, for twice the payload. That is the "arms got cheaper" story actually being true, with the vendor's own published numbers on both ends.

How it was achieved is visible in the product change. Trossen stopped selling two different arms and started selling one. The old kit paired a large follower with a small leader; the new one uses identical WidowX AI arms in both roles, with only the end fitting differing. Building one arm instead of two, in higher volume, is the ordinary way a manufacturer takes cost out.

Why it matters is that this is the rig most published bimanual imitation-learning work runs on, so its price is the entry fee to reproducing that work. Six thousand dollars came off that fee, and the resulting arms lift twice as much.

What it still cannot do is compete on precision. One millimetre of repeatability is ten times worse than the industrial arms in section 2, and it is the number that decides whether a learned policy can be replayed reliably. It also still has no joint torque sensing, so contact-rich work needs a separate wrist sensor.

Status: Shipping, average production time two to three weeks, published by the vendor.

One more thing on that page is worth reading as a hardware fact about learning. Trossen sells an optional workstation with the kit — an NVIDIA RTX 5090 and an Intel Core Ultra 9 — at $10,995.95. The training computer now costs half as much as the robot.

3.5 Open-source torque control arrived: OpenArm#

This is the most significant new open hardware in the period, and it is the one with the licence trap in it.

Before, an arm you could back-drive by hand and command in torque rather than position meant a Franka or a Kinova, from a vendor that publishes no price, on a procurement timescale measured in months. Torque control is what stiffness and force control needs underneath it, so this was a hard floor on who could do contact-rich work.

What changed is OpenArm, from Enactic. The documentation gives 7 degrees of freedom per arm, a nominal payload of "4.1 kg (held for 1 minute in the worst posture)" and a peak payload of "6.0 kg (3 s move

• 1 s hold in the worst posture)", quasi-direct-drive backdrivable joints, a CAN-FD control bus, an aluminium and stainless steel structure, and a compact parallel gripper with an in-hand camera. Note how the payload figures are qualified: the vendor states the posture and the duration, which is better practice than most of the industry manages and is exactly the habit the gripper document asks you to look for.

It is also buyable without building it. WowRobo lists the OpenArm V1.1 at $5,400 and OpenArm 2 as a bimanual set at $6,500 and $7,300.

How it was achieved is the quasi-direct-drive actuator. A quasi-direct-drive joint uses a motor with a very low gear ratio, so the motor's own torque is felt almost directly at the joint and the joint can be pushed back by hand. Those actuators became a commodity part, and Enactic maintains a repository of notes on the DAMIAO actuators it uses. The arm is a frame around bought-in actuators, which is why it costs a fifth of what a comparable closed arm does.

Why it matters is that force control stopped requiring a purchase-order conversation. A 7-degree-of-freedom backdrivable arm at $5,400 puts impedance control, human-robot contact and compliant assembly inside a normal project budget.

What it still cannot do is be treated as permissively licensed, and this is the trap. The software repositories — openarm, openarm_ros2, openarm_can, openarm_description and the rest — are Apache-2.0. But openarm_hardware, which holds the CAD data and the manufacturing information, is CERN-OHL-S-2.0. The S stands for strongly reciprocal: if you make and distribute a product based on those designs, you must make your own design files available under the same licence. That is the hardware equivalent of the GPL, and it is a completely different obligation from the Apache-2.0 on the software in the same organisation. Reading the GitHub badge on the main repository and concluding the hardware is Apache-2.0 would be a real and expensive mistake.

Status: Shipping, through a third-party seller.

3.6 Whole robots, not just arms#

Three more things became buyable that are not arms but are what people actually put arms on.

XLeRobot is a dual-arm mobile home robot built on SO-101-derived arms, Apache-2.0, at version 0.3.0 since 30 August 2025, and the repository states a basic configuration cost of about $660, with $30 for a stereo head camera, $79 for a Raspberry Pi and $220 for a RealSense depth camera as options. WowRobo sells kits at $239, $349 and $579. The repository's own honesty about what it costs you is worth quoting: assembly requires "at least a day" and familiarity with programming.

Reachy 2 from Pollen Robotics, now "Part of Hugging Face", has two 7-degree-of-freedom arms with about 3 kg of payload each, and the software is Apache-2.0. The page describes the hardware as open but does not name the hardware licence, so this document does not state one. No price is published on that page.

Reachy Mini is the small desk robot from the same team, at $399 for the lite version and $499 for the wireless version with a Raspberry Pi 4 and a battery, with a quoted delivery time of about ninety days. It has no arms and no gripper, which is worth saying plainly because it is frequently mentioned in the same breath as the arms and it cannot manipulate anything.

Status: XLeRobot and Reachy Mini, Shipping. Reachy 2, Shipping, price on application.

4. Humanoid arms and hands#

Humanoids are outside this repository's subject, which is arms bolted to tables. They appear here for one reason: the money that went into them paid for arms and hands that you can now buy, and those arms and hands work perfectly well on a table. This section covers only that part, and ignores legs, balance and locomotion entirely.

4.1 What is actually buyable, and at what price#

Before, a humanoid was something you saw in a video. The arms inside it were not sold, no price was published, and the specification sheets did not exist.

What changed is that Unitree and AgiBot both opened storefronts with prices in United States dollars. Read the following table as: the robot, what one arm can lift, and what the vendor charges. Every figure is from the vendor's own page, fetched in September 2026, and a dash means the vendor publishes nothing.

Three things in that table matter more than the prices.

The arm payload is the number to read, and it is small. A humanoid that carries 25 kg carries it with both arms and a simple end fitting. Per arm, with a hand on the end, the figures are 2 kg on a G1, about 2 kg on an AgiBot A2 Ultra, 7 kg rated on a Unitree H2 and 8.2 kg on a 1X NEO. Those are the numbers that decide what a manipulation task can involve, and they sit in the same range as the research arms in section 3.

A cheap humanoid may not be programmable. Unitree's own store page notes that the $13,500 G1 "does not support secondary development". Research use requires the EDU edition, which is quoted rather than priced. Buying the headline price and then discovering you cannot write code for it is a real failure mode.

1X NEO is an order page, not a product. The order page takes a $200 deposit described as "Fully Refundable" and says "US Deliveries start 2026". A refundable deposit against a year is the definition of Announced with a date, and this document does not report it as shipping.

4.2 Two arms you can buy without the robot#

The genuinely useful development for anyone working on manipulation is that arms started being sold on their own.

Unitree's store lists the R1-7a, a seven-axis arm, at $1,650, with a student price of $899 shown but unavailable when the store was read. The listing was published on 4 September 2026, which makes it three weeks old at the time of writing. The same store has sold the Z1 arm since 2022 at $15,999.

That contrast is the whole story of this section in two rows of one shop's catalogue. A seven-axis arm from the same vendor went from $15,999 to $1,650 in four years. The two are not the same product — the Z1 is a heavier, older, higher-payload design and Unitree publishes no payload for the R1-7a at all — so this is not a like-for-like price fall and should not be reported as one. What it is, is evidence that the humanoid actuator supply chain now produces an arm at a price the arm industry could not previously reach.

How it was achieved is the actuator again, and the numbers from the humanoid makers say so directly. Figure's own account of ramping production, published on 29 April 2026, reports building "9,000+ actuators" across more than ten distinct types in the course of building its robots. Inspire Robots announced on 8 September 2026 that it had delivered 10,000 dexterous hands cumulatively, a five-fold increase in shipment volume, on an actuator capacity of 100,000 units a year. When a company builds actuators at that rate for one product, the marginal cost of putting six of them in a frame and selling it as an arm collapses.

Why it matters is that the cheapest seven-axis arm is now cheaper than the cheapest six-axis industrial arm by a factor of five.

What it still cannot do is be stiff, and Unitree says so itself. The R1-7a's product page gives a 2 kg payload, a 650 mm reach, a 5.5 kg arm and both force feedback and collision detection — and then adds a footnote that is more useful than the rest of the page: "Given the relatively low reduction ratio adopted for each joint, the overall positional-control stiffness of the robot arm is limited." A low reduction ratio is what makes a joint back-drivable and cheap, and it is also what makes the arm soft under load. That sentence is the trade, stated by the vendor, and it is the single most honest line on any arm page in this document.

It also tells you what kind of force sensing you are buying. A joint with a low reduction ratio and no output-side sensor estimates its torque from motor current. That works, and it is not the same thing as the strain-gauge joint torque sensors in a Franka or a KUKA LBR iiwa — see section 9.

Status: Shipping, for both.

4.3 The hands: separately purchasable at last, but not from everyone#

Before, a multi-finger hand meant Shadow Robot or Wonik Allegro, neither of which publishes a price, on a procurement cycle to match. The gripper document's section on multi-finger hands has their specifications and records that neither publishes a price. That is still true in September 2026.

What changed is AgiBot. Its store lists the OmniHand 2025 at $4,420 and the same hand with tactile sensing at $5,360 — so the tactile option is a $940 line item you can read off a web page. The OmniHand O12 product page describes it explicitly as "offered as a standalone accessory compatible with multiple robot platforms" and gives 12 active degrees of freedom out of 19 total, five fingers, a typical maximum fingertip force of 20 N, a load of 35 kg with the palm up and 5 kg with the palm down, a weight of 750 g or less, and tactile sensing comprising three-axis fingertip force and one-axis palm force over "150+ tactile points" at a resolution of 0.01 N.

Unitree's hands are the counter-example and the disappointment. The Dex3-1 has 7 degrees of freedom across three fingers, 33 tactile sensors covering 10 g to 2500 g, a maximum held weight of 500 g and a mass of 710 g. The Dex5-1 has 20 degrees of freedom of which 16 are active, five fingers, a fingertip grip force of 10 N, a load of 3.5 kg with the palm down and 4.5 kg with the palm to the side, a fingertip accuracy of ±1 mm and a mass of 1100 g. The tactile version, the Dex5-1P, carries 94 tactile sensors per hand; the base Dex5-1 has none. And neither appears as a purchasable item anywhere in Unitree's store. They are configuration options on an EDU humanoid, quoted by email. You cannot buy a Unitree hand.

Why the AgiBot listing matters is that it is the first time a five-finger hand with fingertip force sensing has had a public price at all, let alone one under six thousand dollars. Against a Shadow Dexterous Hand, which the gripper document describes as quote-only, that is a different kind of purchase.

What it still cannot do is lift anything while being dexterous. The pattern is consistent across every hand in this section: Dex3-1 at 500 g, Dex5-1 at 3.5 kg palm-down, OmniHand at 5 kg palm-down. The 15 kg and 23 kg payload figures that appear in humanoid marketing belong to arms with simple grippers on the end, not to arms with hands on the end. Nobody sells a hand that is both dexterous and strong, and this is the single most reliable limitation in the whole category.

Two smaller cautions about hand specifications, because they are systematically misreported. First, actuated degrees of freedom, total degrees of freedom and joint count are three different numbers, and vendors quote whichever is largest. Unitree's "20 Degrees of freedom (16 active+4)" and AgiBot's "12 active DoF, 19 total" are both honest because they give the qualifier; a secondary source that drops the qualifier is not. Second, tactile sensing is frequently an upsell rather than a feature — the Dex5-1P and the $940 AgiBot option are both separate products from the hands they resemble.

Status: AgiBot OmniHand, Shipping with a published price. Unitree Dex3-1 and Dex5-1, Shipping but bundled only, no published price. Shadow and Allegro, Shipping, quote only.

4.4 Which humanoids have actually been built, and in what numbers#

This is the part the announcements are worst at, so it is worth separating the verified numbers from the rest.

Figure publishes the most detail of anyone. Its production account of 29 April 2026 reports "over 350 of our third generation humanoid robots" delivered, a rate that went "from 1 Figure 03 per day to 1 per hour" in under 120 days, more than 9,000 actuators, over 500 battery packs, an end-of-line first-pass yield above 80 per cent and a battery line at 99.3 per cent. Its earlier announcement of the BotQ factory of 15 March 2025 claims capacity for 12,000 robots a year and a goal of 100,000 over four years. Figure 03 itself, announced 9 October 2025, publishes that its fingertip sensors "detect forces as small as 3 grams" and that each hand contains an embedded palm camera. It does not publish the hand's degrees of freedom, and this document does not state one.

Three hundred and fifty robots is a real number and it is also a pilot fleet. Against it, Inspire Robots delivered 10,000 hands. The category that reached volume manufacturing is the hand, not the humanoid.

Several well-known robots publish very little. Boston Dynamics' Atlas gives 56 total degrees of freedom, a 2.3 m reach, an instantaneous weight capacity of 50 kg and a sustained capacity of 30 kg, and states that it has "tactile and 360° camera view" — but publishes no per-arm degrees of freedom, no hand specification, no taxel count and no price. Agility's Digit 5 gives 23 kg of payload and 2.2 m of reach and states on the page itself that "Digit 5 is in development and the features, specifications, capabilities, and designs shown are subject to change". Apptronik's Apollo 2 publishes no numbers at all — no payload, no degrees of freedom, no price, no date.

Three vendors could not be checked and this document quotes nothing for them. Tesla's Optimus page returns HTTP 403 to any automated request, so the widely repeated 22-degree-of-freedom hand figure for Optimus Gen 3 is not verified here. Be careful with that number in particular, because 1X publishes 22 degrees of freedom per hand for NEO, and the two claims are easy to confuse. Fourier's site is a client-side application that serves no content to a fetch, and its alternative domain fails its TLS handshake. RobotEra's site, which would carry the XHand specification, is client-rendered in the same way. Those three products exist; their numbers are not reported here because none could be obtained from a source that resolves.

Status: Figure 03, Shipping to Figure's own customers and not for sale. Atlas, Demonstrated, with Boston Dynamics describing early field testing at Hyundai and publishing no dates or unit counts. Digit 5, Announced. Apollo 2, Demonstrated. Optimus, Fourier and RobotEra, unverified.

5. Grippers and multi-finger hands#

Grippers and the hardware around them is the catalogue, and it is current to September 2026. This section records only what moved.

5.1 Robotiq now maintains its own ROS 2 driver#

Before, the most widely used driver for the most widely used collaborative gripper was a community project. The gripper document records the position exactly: the README of PickNikRobotics/ros2_robotiq_gripper says "this is not sponsored or maintained by Robotiq", and its released-package table covered Humble, Iron and Rolling but not Jazzy, which is the long-term release most cells run on.

What changed is that on 26 August 2026 Robotiq released official ROS 2 packages. They live at github.com/robotiq/ros under the BSD-3-Clause licence, build Humble, Jazzy and Lyrical from a single branch with Rolling on non-blocking continuous integration, and run the gripper at 200 Hz. Robotiq describes them as a drop-in successor to the community driver. The repository was last pushed on 23 September 2026, the day this document was written.

How it came about is the ordinary consequence of a product becoming infrastructure. A vendor tolerates a community driver until enough customers depend on it, and then the support burden of not owning it exceeds the cost of owning it.

Why it matters is that the Jazzy gap is closed, and it closes the specific warning in the gripper document's driver table. If you deferred a Robotiq gripper because its driver did not target your distribution, that reason has expired.

What it still cannot do is change the licence position elsewhere. Schunk's ROS 2 drivers remain GPL-3.0 and GelSight's software development kit remains GPL-3.0, and both of those are still the live licensing traps in that table.

Status: Shipping.

The driver is one of three pieces. Robotiq groups them under the name Contact Core, its software layer for what it calls physical artificial intelligence: the ROS 2 package, a standalone C++ software development kit, and a set of Isaac Sim simulation assets. The assets were published on 22 September 2026, one day before this document was written, with variants for both the PhysX and the Newton physics backends. Newton is built on MuJoCo-Warp and, according to Robotiq, "models a closed loop natively, as an equality constraint rather than as a tree plus a patch", which is the thing that makes an underactuated gripper hard to simulate correctly. A gripper vendor maintaining its own simulation models is new behaviour, and it is the useful kind, because the gripper document's note on mujoco_menagerie describes how awkward third-party gripper models are to license correctly.

The assets repository is the one to read carefully. github.com/robotiq/isaacsim_assets is BSD-3-Clause with a carve-out: parts of the PhysX 2F-85 asset were adapted from NVIDIA's Isaac Sim and SimReady libraries and remain under CC BY 4.0, with separate licence files shipped beside the affected files. GitHub's own licence detection reports the repository as unclassified for exactly that reason. A mixed-licence repository is not unusual and it is not a problem, but it is not what a one-word badge would tell you.

Robotiq says the next step, embedded intelligence in the fingertip itself, is due in the fourth quarter of 2026. That is Announced with a date.

One adjacent event from the same day belongs here because it affects the software underneath every arm in this repository. On 23 September 2026 Qualcomm announced that it will acquire PickNik Robotics, the maintainers of MoveIt, and stated that MoveIt will remain open source. PickNik also wrote the community Robotiq driver that Robotiq has now replaced. MoveIt 2 is the motion planning framework this repository uses, so its ownership is worth knowing about. The acquisition is Announced; nothing has completed.

5.2 Tactile fingertips became a gripper accessory#

Before, adding touch sensing to a production gripper meant a research sensor, a mount you designed yourself and a driver you integrated yourself. The gripper document's section on gripper sensors describes that world: GelSight Mini as a camera behind a consumable gel, Contactile's PapillArray as a research instrument, and nothing that a factory would buy as a spare part.

What changed is Robotiq's TSF-85 tactile sensor fingertips, launched in January 2026. They replace the standard fingertips on the 2F-85 and 2F-140 grippers. Each one carries 28 taxels — individual pressure-sensing elements — in a four-by-seven grid, samples at 1000 Hz, covers a force range of 0 to 225 N, includes a three-axis inertial measurement unit for vibration, measures 28.4 mm by 44 mm, and is stated as "Tested to over 2 million cycles".

How it was achieved is packaging rather than physics. A taxel array at this resolution is not new; putting it in the exact shape of an existing fingertip, on a sensor rated for two million cycles, is the part that took a manufacturer.

Why it matters is the cycle rating. The GelSight Mini gel is rated for 1,000 coin presses and is a consumable; two million cycles is the same order as the gripper's own service interval, which means a tactile fingertip can now live in a production cell rather than on a bench.

What it still cannot do is be priced. Robotiq publishes no price and the page offers a quote request, so this is not a part you can budget for from public information.

Status: Shipping, price on application.

5.3 Multi-finger hands became genuinely affordable#

Before, a research hand cost what Shadow or Wonik Allegro quoted, and neither publishes a number. The practical floor for anything with five independently controlled fingers was tens of thousands.

What changed is a spread of options across three orders of magnitude, all of them with published prices, all of them reachable in September 2026. Read the table as: the hand, how many degrees of freedom the maker claims and of what kind, what it costs, and who publishes that cost.

How it was achieved differs at each end of that table and the difference decides what you get. The three cheapest are printed frames around hobby servos, with the design given away and the assembly given to you. The AgiBot and Seed Robotics hands are industrially manufactured products riding on the actuator volumes described in section 4.2. The two at the bottom are low-volume instruments with tendon drives and sensor counts the cheap hands do not attempt.

Why it matters is that the entry price for dexterous manipulation research fell from a procurement exercise to a purchase order, and in the AmazingHand's case to less than the cost of the arm it bolts onto. The cheapest five-finger hand with a price its own maker publishes is now the AgiBot OmniHand at $4,420, with Seed Robotics' RH8D at €5,599 the cheapest from a European vendor.

What they still cannot do is hold anything. None of the cheap hands publishes a payload, and the ones that do publish are small: an AgiBot OmniHand manages 5 kg palm-down and a Unitree Dex3-1 manages 500 g. A cheap five-finger hand is a research instrument for studying grasp and in-hand motion, not a way to pick things up. If picking things up is the job, a two-finger gripper is still the right answer and it is not close.

Four things in that table are widely misreported and are worth stating plainly.

Inspire is not the budget option. It is frequently described as the cheap Chinese hand, and the only Inspire prices that can actually be read run from $20,599.99 to $25,399.99. Those come from an export storefront whose vendor field names Beijing Inspire Robots and whose returns address is an inspire-robots.com one, but Inspire's own corporate site does not link to that store and publishes no prices at all. Treat the figures as the best available and not as vendor-confirmed.

LEAP Hand v2's "$200" is not the cost of the hand. The project's own parts page totals $559.48, of which $249.95 is servos and $309.53 is everything else, including $73.67 of fasteners. The $200 figure is roughly the motor and driver share. A LEAP Hand v2 Advanced at 21 degrees of freedom is quoted at $3,000 with its files to be released on paper acceptance, which is Announced rather than shipping.

The Allegro Hand V5 is a three-finger hand. Wonik's own specification is three fingers of three joints each, giving nine degrees of freedom, at 1,050 g — not the four-finger sixteen-degree-of-freedom device most descriptions assume, which is the V5 Plus. Wonik launched the Allegro Hand V6 F on 7 September 2026 with five fingers, 20 active degrees of freedom and a mass of 1,150 g, and still publishes no price. One wording change across that range is worth noticing because it runs the wrong way: the V4 page describes "sixteen independent torque-controlled joints", and the V5 and V5 Plus pages describe current-controlled joints. That is the same substitution section 9 describes for arms, appearing inside a single product line.

Unitree's Dex5-S is press reporting, not a product listing. A 22-degree-of- freedom hand at ¥39,900 was reported on 22 September 2026. Unitree's own product page for it returns HTTP 404 and no Dex hand of any kind appears in the shop's product feed. Do not treat ¥39,900 as a vendor price. Status: Announced.

Three licence cautions in that table, because hardware licences are misdescribed more often than they are read.

AmazingHand's repository is Apache-2.0, read from the repository itself in September 2026. That is a software licence applied to a hardware project, which is common and legally awkward, because Apache-2.0 talks about source code and object code and says nothing about design files. It is permissive in intent and you should read it as such, but it is not a hardware licence.

RUKA's repository is MIT, from the project's code, with the same caveat.

LEAP Hand v2's CAD is not simply downloadable. The site offers a CAD request form rather than a file, and states no licence on the page. Its predecessor's driver, as the gripper document records, is CC BY-NC 4.0, which forbids commercial use. A $200 hand you cannot use in a product is still a $200 hand, but you should know which one you have bought.

5.4 The 64-fold downcost nobody noticed: Pinc'Open#

Before, Reachy 2's gripper was the "Pincette": machined metal driven by a Dynamixel XM430-W210 servo. Pollen Robotics states its own bill of materials at about €1,700 per unit at low volume.

What changed is that Pollen open-sourced a reimplementation. Pinc'Open replaces the machined metal with printed parts and the Dynamixel with a Feetech STS3215, and is designed to bolt onto an SO-ARM100 or SO-101. Its own bill of materials is €26.46: €14.49 for the servo, €6.29 for the servo driver board, and €2.83 for all eleven printed parts together. On an SO-101 the marginal cost is about €12, because the arm already has a servo at that joint.

How it was achieved is the same substitution that produced the cheap arms in section 3, applied to the end effector: a €14 serial-bus servo in place of a €100-class Dynamixel, and printed plastic in place of machined aluminium.

Why it matters is the ratio. The same function, from the same company, went from about €1,700 to about €26 — a factor of about sixty-four — with the loss in precision and durability being exactly what you would expect and exactly what the application can usually absorb.

What it still cannot do is be combined carelessly. Pinc'Open is licensed CC BY-SA 4.0, read from the repository's own licence file. That is a copyleft licence: derivatives of those parts must be released under the same terms. The SO-101 it bolts onto is Apache-2.0. Bolting the two together produces an assembly under two different hardware licences with different obligations, and several downstream projects describe the whole stack as Apache-2.0, which is wrong about the gripper.

Status: Shipping, as files.

5.5 New gripper technology, and what is only a paper#

Four things shipped in the period that are not simply cheaper versions of existing products.

Schmalz launched the FSGA-78 CF bellows suction cup on 15 May 2026, with a very soft silicone sealing lip that Schmalz describes as working by a "clingfish effect", conforming to rough-sawn timber and textured surfaces that defeat an ordinary cup. That attacks the exact failure mode the gripper document names as suction's first limitation. No price is published.

Schmalz also launched mGrip Basic with an SRCU Basic control module on 2 June 2026, a cheaper tier of the line it acquired from Soft Robotics: up to 10 kg, a workpiece width up to 300 mm, 120 picks per minute, IP69K, with a 24 V control module running on 3.0 to 6.0 bar. Schmalz says it "lowers the entry hurdle" and publishes no price. Status: Shipping.

Toyota released YUBI on 11 May 2026, a teleoperation glove with a matching gripper and mounting flanges for Franka, OpenArm and the Unitree G1. It is worth naming here for its licensing as much as its function: the hardware repository Toyota/yubi-hw is CERN-OHL-W-2.0, the weakly reciprocal variant, while the software sits in a different GitHub organisation under Apache-2.0. Its bill of materials contains thirty-nine lines and no prices at all, and several parts are specified as machined aluminium, so the real cost needs quotes. Status: Shipping, as files.

Two things are frequently described as arriving and have not.

Electroadhesion — holding an object by electrostatic attraction rather than friction or vacuum — is research, not product. Two recent papers exist, a review in npj Robotics on 3 September 2025 and a paper on electroadhesive clutches on 30 March 2026. The company that commercialised the idea, Grabit, still has a live website whose most recent news item is dated 19 June 2018. Status: Demonstrated.

Gecko-adhesion grippers were announced and their shipping status is not stated. geCKo Materials showed four products on 29 October 2025, including a gripper for smooth surfaces such as solar panels and glass and a semiconductor wafer-handling tool demonstrated at 5.4 g of repeated acceleration on a FANUC arm. No prices were given and the announcement does not say the products ship. Status: Announced. OnRobot's own Gecko Gripper, by contrast, is still in its catalogue and has been for years.

5.6 Two things that are repeated and are wrong#

The Fin Ray patent has not been shown to have expired. The compliant finger that bends towards a contact, used by every soft two-finger gripper, is usually said to be freely copyable now. This document could not verify that. Every patent database refused an automated request — Google Patents returned 503, Espacenet and the European Patent Office register returned 403, and the World Intellectual Property Organization returned 404. The best candidate for the founding patent, EP1203640A2 by Leif Kniese with a priority date of 1 November 2000, shows no recorded grant in the one index that answered. A separate Festo patent, EP2735408B1, was filed on 27 November 2012 and granted on 20 April 2016, which by the twenty-year rule runs to 2032 at the outside — that is arithmetic from a verified filing date, not a reading of legal status. Separately, "Fin Ray Effect®" is used as a registered trademark, and a trademark does not expire on a patent's schedule, which is why the open projects all say "fin-ray-inspired" instead. Every claim that the patent has expired that this document checked was unsourced.

Shadow Robot has not gone into administration. The claim circulates. Companies House records company number 03308007 as Active, with share allotments filed in June 2025, May 2026 and June 2026, and no insolvency, liquidation or strike-off filing of any kind.

5.7 One older market event worth not repeating#

Soft Robotics Inc divested its gripper business to the Schmalz Group on 6 August 2024, and its mGrip line is now a Schmalz product. That is already recorded in the gripper document's soft gripper section and nothing has changed since. It is repeated here only because comparison articles listing Soft Robotics as a gripper vendor are still circulating, and they are more than two years out of date.

6. Tactile and force sensing#

This moved further and faster than anything else in this document, and it moved mostly by getting cheap rather than by getting better.

6.1 The six-axis force-torque price floor did not move#

A force-torque sensor at the wrist measures the six numbers that describe how the tool is being pushed and twisted: three forces and three torques. The gripper document explains what you do with one and records the Robotiq FT 300-S specification in detail.

Before, the cheapest six-axis sensor with a price its own maker published was Bota Systems' MiniONE at CHF 3,045.

What changed is nothing. Bota's own shop still lists the MiniONE Gen A at CHF 3,045.00, with its SensONE starting at CHF 4,882.50 and its Rokubi at CHF 5,407.50. The only new published figure anywhere is UFACTORY's, which lists a six-axis force-torque sensor at $3,000, marked down from $4,000 — and publishes no force range, no resolution and no sample rate to go with it. That is the trade in that listing: you get a price and you do not get a specification. ATI, Robotiq, OnRobot and Schunk publish nothing at all, exactly as they did in 2020.

So the honest answer to "did force sensing get cheaper" is no, not at the level anyone publishes. Every sub-thousand-dollar claim this document checked resolved to a marketplace listing with no verifiable specification, or to a research prototype.

What did change is the top of the range and the ownership of it. ATI Industrial Automation has been owned by Novanta since a $172 million acquisition announced on 19 July 2021, and now trades at ati.novanta.com. It added two sensors in 2026, and one of them is explicitly aimed at this document's subject: the ATI Varo, for humanoids, rated at ±3000 N in x and y, ±6000 N in z and ±120 Nm, with a resolution of 0.29 N, running at 8 kHz with under a millisecond of latency at 220 g. A companion sensor for robotic surgery, the Sano74, resolves 0.015 N.

Set that against the sensor most collaborative cells actually use. The Robotiq FT 300-S product sheet is footer-dated February 2021 and its data output rate is 100 Hz. Bota runs at 3800 Hz and ATI at 8 kHz. A 38-fold difference in sampling rate between the common choice and the high end is the real gap in this category, and it is not a price gap.

There is one open-source option and it is worth knowing about for the right reasons. CoinFT, from Stanford, publishes a complete bill of materials totalling $10.93 — two printed circuit boards with comb-shaped electrodes separated by cast silicone pillars, with one microcontroller recovering six axes from a single pair of sensing electrodes by switching modes. Its measured performance is a root-mean-square error of 0.16 N over a 0 to 14 N normal range, at 360 Hz sampling with a mechanical bandwidth of about 97 Hz. Two things stop it being the answer to the price floor. The mechanical bandwidth is an order of magnitude below an industrial sensor's, and the project is licensed CC BY-NC-SA 4.0, which forbids commercial use without a separate licence from Stanford. "Open source, therefore free to build a product on" is wrong here, and it is wrong in the same way for several projects in this document.

Why any of this matters is budgeting. If your plan assumed that force sensing would follow arms down in price, it has not, and there is no published evidence that it is about to.

What might change it is volume rather than design. Link-Touch, a Chinese force sensor maker, raised over 100 million renminbi in April 2026 from investors including AgiBot and Galbot, and expects to ship roughly 400,000 joint force sensors and 30,000 to 40,000 six-axis force sensors in 2026, against a long-term target of a million and 200,000 a year. Those are the volumes that break a price floor. No published price has moved yet.

Status: Bota, ATI, Robotiq, OnRobot and UFACTORY, all Shipping. CoinFT, Demonstrated, as files.

6.2 Touch sensing became something you buy for tens of dollars#

Before, a tactile sensor was a research instrument. GelSight Mini at $510, with replacement gels at $57, was the only tactile product in the whole gripper document with a published price. Everything else was a paper with a repository attached. That has not changed at GelSight: its own store still lists the Mini system at $510.00 and the robotics package at $560.00, the same product since 2022, and its 2025 and 2026 announcements were all about surface metrology rather than robots.

What changed is a manufacturing ecosystem around the open research designs. Read the table as: what the thing is, what it costs, and where the price comes from.

How it was achieved is the same mechanism as the arm kits in section 3.2. AnySkin and ReSkin are open designs from academic groups, published under permissive licences — AnySkin's repository is MIT, and so is eFlesh, its successor, which lets you print a tactile sensor in an arbitrary shape from a printer, magnets costing under $5 and the same magnetometer board — and a manufacturer read the files and made them properly. Magnetic skin is also intrinsically cheap: it is a magnetised elastomer over a board of magnetometers, and magnetometers are a commodity part made in the billions for phones.

There is a second, quieter reason it got cheap, and AnySkin's own paper states it: the sensing electronics are decoupled from the sensing surface. The board stays on the robot and the skin slips over it like a phone case, so the part that wears out is the cheap part. That is a design decision rather than a manufacturing one, and it is why a replacement costs $48 rather than the price of the sensor.

Why it matters is that touch stopped being a project. A $48 skin on a $122 arm is a tactile robot for under two hundred dollars, and that was not possible in 2024 at any price a person would pay from their own pocket.

What it still cannot do is any of the things the gripper document already lists. No manufacturer in this category publishes a slip-detection latency, which is the number you would most want. Magnetic skins drift with temperature and need re-zeroing. Gels wear out. Five magnetometers give fifteen channels, which is force and slip rather than shape or texture.

And there is still no standard for tactile data, so a policy trained on one sensor does not transfer to another — which is exactly the gap that the tactile foundation models described in the mechanism document are trying to close, and they have not closed it yet. That gap now has a measured size, and it is the most useful number in this section for anyone about to buy. A September 2026 study reports that a frozen classifier built on Meta's Sparsh tactile model scores 6.86 per cent on a sensor it was not trained on, rising to 87.09 per cent once 10 per cent of the target sensor's data is labelled. Buying the wrong tactile sensor does not cost you some accuracy. It costs you nearly all of it, and the recovery is a labelling exercise. Check which sensors a model was trained on before you buy the sensor.

Status: Shipping, all of them.

6.3 Tactile sensing arrived inside the hands#

The other route to touch is to buy a hand that already has it, and that became possible in the period.

The specifications the vendors publish, for the hands in section 4.3 and section 5.3. Read this as: whose hand, how much of it senses touch, and what the smallest force it resolves is.

The pattern worth extracting is that the sensitivity figures are now well below what a wrist force-torque sensor can see. The gripper document's Robotiq FT 300-S recommends a 1 N threshold for detecting contact. A fingertip claiming 0.01 N is two orders of magnitude finer, and it is finer because it is in the right place: a wrist sensor reads the whole load below it, and a fingertip reads only the contact.

What it still cannot do is come without the hand. Every entry in that table is a sensor you get by buying a complete hand, most of them at four or five figures, and the tactile version is often a separate more expensive product — AgiBot charges $940 for the option and Unitree's base Dex5-1 has no tactile sensing at all. If you want touch on a gripper you already own, the answer is section 6.2 or the Robotiq fingertips, not a hand.

6.4 The announcement that never became a product: Meta's Digit 360#

This one is included because it is the clearest example in the whole document of why the maturity labels exist.

Before, Meta's DIGIT was the standard cheap vision-based tactile sensor in research, and it is now sold by GelSight at $355.

What was announced is Digit 360, on 31 October 2024. Meta's own words: it "captures forces as small as 1 millinewton", carries "over 8 million taxels" and "over 18 sensing features" including vibration, heat and odour, with an on-device accelerator. Its product page adds spatial detail "as fine as 7 microns". And the announcement said plainly: "GelSight Inc. will manufacture and distribute Digit 360, available for purchase in 2025."

What happened is that it was never sold. GelSight's store has no Digit 360 product, and a search of GelSight's own store for it returns "No products were found". The only distribution route, a free research programme, still shows proposals due on 20 January 2025 and acceptance notification "In progress", with delivery "To be confirmed upon acceptance" — twenty months later. The site footer still reads 2024. The companion announcement, a Wonik Allegro Hand integrated with Meta's Digit Plexus connector and "slated for 2025 launch", also did not happen: Wonik's current hands use capacitive pressure sensors and no Meta component appears on any of their product pages.

Why it matters is not that Meta failed to ship something. It is that between October 2024 and today, a great deal of writing treated Digit 360's specification as a thing you could buy, and the correct label for it was never better than Announced with a date.

What it still cannot do, even for the researchers who did receive units, is underpin a product. Every piece of Meta tactile hardware and every Meta tactile model — digit360, digit-design, digit-plexus, sparsh — is licensed CC BY-NC 4.0, read from each repository's own licence file. GitHub's licence detection reports all of them as unclassified, which is why they are so often described as open source. They are published, and they are not usable commercially. Only Meta's TACTO simulator is permissively licensed, under MIT.

Status: Demonstrated. Not buyable at any price.

7. On-robot compute#

What you put on the robot changed more in this period than the robot did, and it changed in a direction nobody predicted in 2024.

7.1 The one verified price cut: Orin Nano Super, December 2024#

Before, the entry-level NVIDIA Jetson developer kit was the Orin Nano at $499, a figure NVIDIA published in March 2023. The step up was the Jetson AGX Orin Developer Kit at $1,999, upgraded to 64 GB of memory at the same price.

What changed is that on 17 December 2024 NVIDIA published, in its own words, a "New reduced price of $249, down from $499" for the Orin Nano, renamed the Orin Nano Super. At the same moment the compute figure went from 40 to 67 sparse INT8 TOPS — that is trillions of eight-bit integer operations per second, counted with the sparsity optimisation applied — and memory bandwidth went from 65 to 102 GB/s, with a new 25 W power mode.

How it was achieved is the part that makes this the most interesting event in the section. It was a software update. The silicon did not change. Existing Orin Nano owners received the performance increase for nothing, by updating JetPack and selecting the new power mode. NVIDIA had been shipping the hardware below its capability and then unlocked it.

Why it matters is that it halved the price of the standard entry-level robot computer, and it is the only NVIDIA Jetson price reduction this document could verify from NVIDIA's own publication.

What it still cannot do is tell you today's price. NVIDIA prints no prices on any current Jetson product page, and both of the stores its own "Shop Now" buttons point at — its marketplace and its Jetson store — failed to respond to repeated automated requests. Every dollar figure in this section is therefore a launch price published in an NVIDIA blog post on a stated date, not a street price today. Given section 7.4, the difference may be large.

Status: Shipping.

7.2 Jetson Thor, and how to read its headline number#

Before, the top of the Jetson range was the AGX Orin at up to 275 sparse INT8 TOPS in a 15 to 60 W envelope.

What changed is Jetson AGX Thor, announced on 25 August 2025 with a developer kit price of $3,499. The T5000 module carries a 2560-core Blackwell graphics processor, a 14-core Arm Neoverse-V3AE processor, 128 GB of 256-bit LPDDR5X memory at 273 GB/s, and a power envelope of 40 to 130 W. NVIDIA states its compute as 2070 TFLOPS, and the footnote on its own specification table gives the condition: FP4, sparse, measured at 130 W. The smaller T4000 gives 1200 FP4 TFLOPS with 64 GB. NVIDIA guarantees availability of the T5000 through August 2035.

How to read the headline is where care is needed, and this is the single most misquoted number in robotics hardware. NVIDIA claims Thor delivers "7.5× the performance and 3.5× the energy efficiency of NVIDIA AGX Orin". That comparison puts Thor's FP4 sparse figure against Orin's INT8 sparse figure. Those are different number formats. FP4 is a four-bit floating-point format and INT8 is an eight-bit integer one, and an operation in one is not an operation in the other. NVIDIA publishes no dense FP4 figure for Thor at all. The 7.5× is a real engineering achievement and it is not a like-for-like multiple, and anyone repeating it without saying so is passing on a marketing number.

The same care applies one level down. NVIDIA publishes both sparse and dense figures for the Orin family, and every headline in circulation is the sparse one: the AGX Orin 64 GB is 275 TOPS sparse and 85 TOPS dense on its tensor cores, and the Orin Nano Super is 67 sparse against 33 dense. If you are sizing a model against a TOPS figure, the dense number is the one that describes what you will get. NVIDIA's own Jetson Orin page currently carries a contradiction on this point, listing the AGX Orin 32 GB at both 241 and 200 sparse TOPS in different tables; a June 2026 NVIDIA blog post resolves it in favour of 241, describing a 20 per cent boost above the original specification.

Why Thor matters is memory rather than arithmetic. 128 GB of unified memory on a robot means a model that previously needed a workstation now fits.

What it still cannot do is feed itself. 273 GB/s against 2070 TFLOPS makes large-language-model decoding bandwidth-bound rather than compute-bound, and NVIDIA's own MLPerf submission of 16 September 2026 shows what that means in practice: 52.33 output tokens per second on a 27-billion-parameter model at NVFP4, with a median time to first token of 247 ms, on a single developer kit at maximum power. It also cannot run on a battery comfortably. Its floor is 40 W against the Orin Nano's 7 W, and on a mobile robot that is a design constraint rather than a detail.

Status: Shipping. NVIDIA stated on 15 July 2026 that "Developers can begin building today using the Jetson AGX Thor developer kit available through channel partners".

7.3 What was announced and has not arrived#

Two Jetson products are Announced with a date and neither is buyable.

The T3000 and T2000 modules, announced on 15 July 2026, give 865 and 400 FP4 TFLOPS, with the T3000 at 32 GB and described as roughly half the size and power of the T5000. NVIDIA states they "are scheduled to become available in Q1 2027" and publishes no price. NVIDIA's stated reason for the smaller parts is worth quoting because it is unusually candid: "Migrating to T3000 helps reduce costs amid high memory prices."

The Jetson Orin Nano 2, announced on 25 August 2026, gives "78 trillion operations per second of AI compute, 8GB of memory and an 8-core Arm CPU", is claimed to deliver "2x the inference performance of Jetson Orin Nano Super", and to consume "40% less power to deliver the same performance as its predecessor" in 15 W mode. It is expected in the first half of 2027 and no price is published. NVIDIA does not state the precision behind the 78 figure, so this document does not either.

The instructive detail is the arithmetic. Going from 67 to 78 is a factor of 1.16, and NVIDIA claims twice the inference performance. The gain comes from memory bandwidth and improved tensor cores, not from arithmetic throughput. NVIDIA's own product is the clearest available demonstration that TOPS is the wrong number to choose a robot computer by.

7.4 The actual story of 2026 is memory prices#

This is the thing a reader planning a budget most needs to know, and it has nothing to do with robots.

Before, a Raspberry Pi 5 with 8 GB of memory cost $80 and was the default cheap robot computer.

What changed is that memory prices rose sharply, and Raspberry Pi documented it in its own words across several dated posts: "an unprecedented rise in the cost of LPDDR4 memory, thanks to competition for memory fab capacity from the AI infrastructure roll-out", and "a seven-fold increase over the last year in the price of the LPDDR4 DRAM used on Raspberry Pi 4 and 5". The published prices moved accordingly: the 2 GB Pi 5 from $50 to $65, the 4 GB from $60 to $110, the 8 GB from $80 to $175, and the 16 GB from $120 to $305. Raspberry Pi introduced a new 1 GB Pi 5 at $45 to have anything cheap left in the range, and calls the pressure "painful but ultimately temporary".

How it happened is the data-centre build-out competing for the same fabrication capacity. It is the clearest case in this document of the robot market being a passenger in somebody else's economy.

Why it matters is immediate and practical. Any board price you remember from before 2026 may be wrong by a factor of two. An 8 GB Pi 5 more than doubled. The same force is visible on the NVIDIA side in the sentence quoted in section 7.3.

What it still cannot do is be predicted. Raspberry Pi says temporary and does not say when.

Status: current, and the reason to re-check every price before you buy.

7.5 The alternatives, and what each publishes#

Read this table as: the part, what its maker claims for compute and in what precision, and whether a price is published anywhere by the maker.

Five things in that table are worth more than the numbers.

Texas Instruments is the only vendor here that publishes a price for a chip. It is also the only one that has shipped nothing new: the whole TDA4 and AM6xA family is on a 16 nanometre process and its pitch is functional safety and backward compatibility rather than throughput. For a robot that has to be certified, that is the point rather than a weakness.

Qualcomm's 700 TOPS headline is half that at INT8 dense, and Qualcomm says so on its own page. This is the same reading problem as Thor's 7.5×, and Qualcomm handles it more honestly by printing both numbers side by side.

Qualcomm's most consequential move is not silicon. It announced on 23 September 2026 that it will acquire PickNik Robotics, the stewards of MoveIt, stating that MoveIt 1 and MoveIt 2 will "remain open, community-driven, and supported across third-party hardware platforms". A chip vendor buying the motion-planning framework is a bid for the layer above the silicon.

Hailo does not state the precision behind its two best-known figures. The 26 TOPS of the Hailo-8 and the 13 of the Hailo-8L are printed without a format, while the Hailo-10H's "40 TOPS of INT4" is qualified. At matched INT8 precision the newer Hailo-10H is 20 TOPS against the older Hailo-8's 26, which is why Raspberry Pi describes the new board's vision performance as "broadly equivalent" to its predecessor. The gain is the 8 GB of memory, not the throughput.

Google Coral is gone without being discontinued. There is no end-of-life notice, and the boilerplate that would carry one has not been used. But coral.ai now redirects wholesale to a Google developer page, the legacy news page's most recent dated release is from July 2021, and the successor — an open-source RISC-V design licensed Apache-2.0 — targets a ten-milliwatt wearable envelope, which is a direction away from robots. Google's own words are that it is "an evolution of Google's edge AI strategy, rather than a direct replacement for Coral". If you have Coral in a design, plan to move.

Two fetching notes belong here because they affect how much of the above could be checked. Raspberry Pi and Hailo both return HTTP 403 to an automated request with browser headers, which is a bot block rather than a dead page, so their figures were read through a renderer instead. And NVIDIA's news host returns HTTP 200 for any path at all, so a 200 there proves nothing on its own.

7.6 What a robot actually runs, as against what it could#

The gap between the top of the range and the deployed fleet is wider than the announcements suggest, and NVIDIA's own customer list shows it. Its June 2026 account of named deployments puts Zipline's delivery drones and SandStar's vending machines on Orin NX, and describes Hexagon and Solomon as integrating Thor into humanoids. An earlier NVIDIA post describes Caterpillar's assistant on Thor as "in development". Even the Unitree G1, at $13,500, specifies only an "8-core high-performance CPU" in its base form, with Orin appearing as an option on the EDU version.

The pattern is that shipping robots run Orin NX and Orin Nano, while Thor is in humanoids, demonstrations and development. That is the honest difference between 2024 and 2026: the top of the range moved a long way and the deployed fleet did not move much at all.

One independent benchmark is worth reporting because it reframes the question, with its disclosure attached. Open Navigation, the company behind the Nav2 navigation stack, published a robotics workload benchmark in July 2026 comparing Jetson Thor, Jetson AGX Orin and an AMD Strix Halo part on a simulated forklift with nine sensors. Open Navigation states that the work was a collaboration with AMD and that the benchmark was executed independently, and you should weigh the results with that in mind. Its finding is that Thor's misses in the control loop came from its Arm cores saturating on trajectory planning — work that cannot be moved to the graphics processor — and that the AGX Orin ran at 93.6 per cent mean processor utilisation and completed three of ten missions. Whether or not the comparison favours the sponsor, the mechanism it identifies is checkable and matters: on a modern ROS 2 workload the binding constraint is single-core processor performance for planning, not graphics-processor TOPS.

7.7 Real-time, which did not arrive on Jetson#

A robot that must guarantee a control loop needs a real-time operating system, meaning one that bounds how long a task can be delayed rather than merely making delays rare.

Before, real-time Linux meant applying the PREEMPT_RT patch set to the kernel yourself, and it had been an out-of-tree patch for about two decades.

What changed is that PREEMPT_RT was merged into mainline Linux in version 6.12, released on 17 November 2024. What remains out of tree is now tiny: the current supplementary patch is a few hundred lines and touches nothing in the scheduler, the locking code or the timer core.

What did not change is Jetson. NVIDIA's own documentation still says that "Real-Time Kernel support is provided with Developer-Preview quality" for the Thor and Orin modules, and has said so since release 35.1. JetPack 7 ships a 6.8 kernel, which predates the 6.12 merge, so Jetson's real-time support is still a patched out-of-tree kernel rather than mainline PREEMPT_RT. Building the display drivers under it requires setting a flag that overrides NVIDIA's own check for the presence of PREEMPT_RT, which tells you how well tested the combination is.

Why this matters is that the compute story and the control story point in opposite directions. You can now put a 2070-teraflop computer on a robot, and the part of the software that has to hit a deadline every millisecond is running on a kernel its vendor labels a developer preview.

One more scheduling fact belongs here because it will bite anyone pairing the two. The current ROS 2 long-term release is Lyrical Luth, released on 22 May 2026 and supported to May 2031. Its tier-one platform is Ubuntu 26.04, and JetPack 7 ships Ubuntu 24.04, which Lyrical supports only as a source build. In practice Jazzy, supported to May 2029, remains the distribution that pairs with a Jetson, and this document could not reconcile that with NVIDIA's own claim of Lyrical support on Ubuntu 24.04 in Isaac ROS 5.0.

8. The category that did not exist: teleoperation hardware#

This section is here because nothing in the brief for this document asked for it, and it is the largest genuinely new hardware category of the period.

Before, if you wanted to collect demonstrations for imitation learning you built your own leader device. GELLO — a printed replica of the robot's kinematics with encoders instead of motors — was a paper and a repository, and everyone who wanted one made one.

What changed is that the leader arm became a product, from several directions at once, in a single eighteen-month stretch.

Enactic sells the OpenArm KER, which it calls a Kinematic Equivalent Replica: a lightweight leader arm that matches the follower's joint layout. WowRobo lists it at $2,599, and the individual encoder unit at $49. The software is Apache-2.0; as with the arm itself, the hardware files sit in a separate repository with a separate licence, so check it.

Franka Robotics lists Franka GELLO and GELLO Duo on its own product page, labelled as prototypes. A major industrial arm vendor productising an academic teleoperation rig is a reasonable marker for when a research idea has finished arriving.

Elephant Robotics sells the myController S570, described as a portable exoskeleton robot controller for data capture, at $1,300 on its own store. Unitree prices a whole teleoperation rig rather than a device: its D1-T arm is sold in a dual-arm teleoperation kit at under $8,500 and a quad-arm kit at under $16,000, both before tax and freight. AgileX lists a Universal Master Arm and the PIKA PRO handheld data-collection device on its store, both without a published price. Trossen sells a teleoperation product called Glide with a Cockpit operator station, also without a published price.

Hugging Face went in the opposite direction and removed the robot from the loop entirely. Grabette, published on 21 July 2026, is "a handheld gripper instrumented with everything needed to reconstruct a manipulation demonstration" — two cameras with distinct jobs, an inertial measurement unit and a gripper — with a bill of materials of about €490, and a companion motorised gripper called Gripette at about €120. It writes "a standard LeRobot dataset on the Hugging Face Hub", storing demonstrations as a camera-local six-degree-of-freedom pose plus a gripper state. Its ancestor is UMI, the universal manipulation interface.

There is also an industrial version of the same idea. Universal Robots' news centre records, on 16 March 2026, a joint launch with Scale AI of the UR AI Trainer, described as generating training data in dedicated training cells. A training cell is a room full of hardware whose only product is demonstrations.

How this category came about is the data constraint that the mechanism document puts first among its five forces. If demonstrations are the binding constraint on what a learned policy can do, then a device that makes demonstrations cheaper is worth more than a better arm, and the market responded to that literally.

Why it matters, and this is the practical point for a reader with a small budget, is that a €490 handheld recorder lets you collect a dataset before you own a robot at all. That inverts the usual order of a robot-learning project.

What it still cannot do is close the gap between a human hand and a robot. A handheld recorder captures a trajectory a person made with human wrists and human compliance, and the robot that has to reproduce it has neither. Whether policies trained this way transfer is an open question, not a settled one, and the learned-methods document is the place to read about it.

Status: OpenArm KER, Grabette, Gripette and the myController S570 are Shipping. Franka GELLO and GELLO Duo are Demonstrated, labelled prototypes by the vendor. The UR AI Trainer is Announced with a date.

9. What did not change#

A document about change earns its keep by being specific about what stayed still, because those are the numbers you can plan around.

Mid-range industrial arm prices did not move. The UFACTORY xArm 6 was $8,399 in June 2024, $8,399 in January 2025 and $8,399 in September 2026. That is the only mid-range arm with a public price series, and it is flat.

The major cobot vendors still publish nothing. Universal Robots, Franka Robotics, Kinova, Doosan, Techman, Elite Robots, JAKA and Dobot were all fetched in September 2026 and none of them publishes a price for any product. This has not changed in years and there is no sign of it changing.

Research hands from the established vendors are still quote-only. Shadow Robot and Wonik Allegro publish no price, exactly as recorded in the gripper document.

Precision at the cheap end is still a millimetre. Trossen's WidowX AI states 1 mm of repeatability. The industrial arms state 0.1 mm. That factor of ten did not close, and it is the number that decides whether a demonstration can be replayed.

Dexterous hands still cannot carry anything. Every hand with more than three fingers in this document carries between 500 g and 5 kg, and the strong ones are the least dexterous. This has been true for a decade.

The bottom of the cheap-arm market stopped falling. The SO-101 bill of materials fell about five per cent in twenty-three months and the servos are now most of the cost, so there is no further saving of the kind that produced the hundred-dollar arm in the first place.

Suction, magnetic and tool-changer hardware produced no event worth recording. Every figure in the corresponding sections of the gripper document is current, and nothing in the period displaced a Schmalz cup or an ATI changer.

Nobody publishes a slip-detection latency. The gripper document noted this gap and it is still a gap. A 1000 Hz sampling rate, which Robotiq's new tactile fingertips also quote, bounds the latency from below and is not the same number.

The six-axis force-torque price floor did not move. Bota's MiniONE is still CHF 3,045, ATI and Robotiq and OnRobot and Schunk still publish nothing, and the only genuinely cheap option is a research design that forbids commercial use. Section 6.1 has the detail.

No arm gained joint torque sensors. A joint torque sensor is a strain gauge on the output side of the gearbox, measuring what the link actually feels. Every arm that ships them in September 2026 — the KUKA LBR iiwa, the Franka Research 3, the Agile Robots Diana 7, the Kinova Gen3, the Doosan M and H series — already shipped them in 2022, and not one of those vendors publishes a price, so no price trend for the capability can be established at all. Universal Robots' September 2026 Gen 7 launch did not add them. The thing most likely to be mistaken for progress here is Unitree going from $15,999 to $1,650 on a force-controllable arm, and that arm estimates torque from motor current on a low-reduction-ratio drive, which its own footnote admits. The component supply is densifying — PaXini sells a joint torque sensor at ¥999 and Link-Touch expects to ship roughly 400,000 joint force sensors in 2026 — but none of that has yet reached a cheap finished arm.

10. How to check an announcement yourself#

The research for this document produced a small set of habits that are worth more than any individual number in it, because they will still work next year.

Find out which channel the vendor sells through, because it tells you whether a price exists. A vendor with a web shop publishes everything; a vendor with a quote form publishes nothing. There is no middle case, and knowing which one you are dealing with saves an hour.

Try the shop's own data feed before reading the page. Many robot vendors run their stores on standard e-commerce platforms that expose a machine-readable product list. Appending /products.json to a Shopify storefront, or /wp-json/wc/store/products to a WooCommerce one, returned the complete priced catalogue for four of the vendors in this document in a single request, including products that the human-readable pages do not list together. This is public information the vendor chose to publish; it is simply published in a second place.

Use dated snapshots for historical prices. The Internet Archive holds the same product page from previous years, and the difference between two snapshots is the only honest way to say whether something got cheaper. Every "it used to cost" claim in this document has a snapshot behind it, and the snapshot has a date printed on it.

For open hardware, find the repository holding the design files and read its own licence file. Do not read the badge on the main repository. OpenArm is the worked example: the software is Apache-2.0 and the CAD is CERN-OHL-S-2.0, which is strongly reciprocal and obliges you to publish your own design files if you distribute a product based on it. The two repositories belong to the same organisation and their licences impose completely different obligations.

Check that a degrees-of-freedom figure carries its qualifier. Actuated, total, and joint count are three different numbers. A vendor that writes "20 degrees of freedom (16 active + 4)" is being honest. A secondary source that writes "20 degrees of freedom" has removed the information.

Check that a payload figure carries its conditions. OpenArm's "4.1 kg (held for 1 minute in the worst posture)" is a usable number. A bare "4 kg" is not, for the reasons the gripper document gives: payload is a static figure and acceleration spends it.

Treat an unreachable page as unverified, not as confirmed. Several vendors in this document serve nothing to an automated request, either by blocking it or by rendering everything in the browser. That is a reason to say a number could not be obtained. It is never a reason to take the number from somebody who resells the product.

Ask what maturity label the announcement deserves. A deposit is not a delivery, a video is not a product, and a page with no date on it is not a plan. Applying the four labels from section 1.1 to an announcement takes about a minute and it is the single most useful minute in this entire subject.

Back to the rest of the frontier area, or across to what is changing in robot manipulation for why hardware moves the way it does. For what any of this hardware can actually hold, see grippers and the hardware around them.