Where manipulation data comes from, and what changed in 2026

A robot manipulation policy is a function from pictures to motor commands, and somebody has to supply the examples it is fitted to. Almost every other question in this field — which architecture, which simulator, which arm — has several workable answers. The question of where the examples come from has, until very recently, had one workable answer: a person drove the robot by hand, one episode at a time, for a week. That is the constraint this document is about, and 2026 is the year several routes round it stopped being papers and started being things you can download.

This is the events document for data. It says what was released, when, by whom, on what terms, and what it can and cannot do. The mechanism behind the events — why the field moves the way it does, and how to judge a new announcement — is in what is changing, which is worth reading first if you want the reasoning rather than the record. Overlap with it is deliberate and small: that document explains why the field moves, and this one records what moved.

Who this is for, and what it is for#

This is written for someone who understands what an imitation-learning demonstration is, and who now has to decide where to get several hundred or several thousand of them. You do not need to have trained a policy. Every term is explained where it first appears.

Three commitments shape it. They are the same commitments the object perception area makes, adapted to a subject where the artefact in question is a pile of recordings rather than a model.

Every item carries a maturity label, because robotics reporting routinely blurs the difference between a thing you can buy, a thing you can download, a thing somebody filmed once, and a press release. Section 2 defines the five labels.

Every licence was read from the project's own licence file or its dataset card, never from a badge, a blog post or a paper. In this area that matters more than anywhere else in the repo, because a dataset licence is as capable of stopping your product as a code licence, and several of the largest and most attractive corpora are not usable commercially. Section 6.5 lists the ones that will catch you out.

Every number says what it was measured under, and in particular what one "demonstration" or one "episode" meant in the dataset that reports it. Section 2.1 explains why that unit is not comparable between datasets, and why almost every league table of dataset sizes you will see is therefore meaningless.

Contents#

1. Why data is the binding constraint
2. How to read the numbers on this page
3. Teleoperation rigs, and the collapse in their price
4. Handheld grippers: collecting without a robot
5. Exoskeletons, gloves and ambient capture
6. The open datasets, and their licences
7. LeRobot, and the tooling that grew round it
8. Human video instead of teleoperation
9. Data scaling laws, and what they actually say
10. Making demonstrations instead of collecting them
11. What changed between May and September 2026
12. What none of this fixes

1. Why data is the binding constraint#

A large language model is fitted to text that already existed. Nobody wrote the internet in order to train one. A manipulation policy is fitted to recordings of a robot being moved through a task, and those recordings did not exist before somebody decided to make them. There is no found corpus. Every episode was paid for in somebody's time.

The size of the gap is easy to state. Open X-Embodiment, the pooled corpus that the field treats as its largest public asset, was assembled by merging sixty existing datasets from 34 laboratories, and its live index now lists 72 member datasets totalling 2,419,193 episodes across 27 distinct robots. A single modern language-model pretraining run consumes text on a scale that is not usefully comparable. The whole of public robot manipulation data, pooled from every laboratory willing to contribute, is a rounding error next to what the adjacent fields consider a starting point.

That shortage is what every development in this document is a response to. There are only five routes round it, and the rest of the document is organised by them.

• Make the robot rig cheaper, so more people collect. That is section 3.
• Remove the robot from collection entirely, and record a person holding an instrumented gripper. That is section 4.
• Record the person's own body instead, with an exoskeleton, a glove or a room full of cameras. That is section 5.
• Pool what exists, publish it, and standardise the format so it can be pooled. That is sections 6 and 7.
• Stop collecting and start generating, either from video of people doing the task, or from a simulator. That is sections 8 and 10.

Section 9 is the question that decides how much any of it is worth: whether more data reliably buys more capability, and at what rate.

2. How to read the numbers on this page#

2.1 What a "demonstration" means, and why the counts do not compare#

Every dataset in this document reports a number of episodes, demonstrations, trajectories or hours. Those units are not the same thing, and even within one unit the underlying object differs enough that arithmetic across datasets is meaningless. Three examples make the point.

In the ALOHA simulation tasks that most people start with, one episode is one continuous recording of a two-armed robot performing one short task from a randomised start, typically around twenty seconds, recorded at 50 Hz with the leader-arm command stored as the action. Fifty of those is a working dataset for that task.

In ABC-130K, released in June 2026, one episode is also a continuous teleoperated recording, but the corpus holds 3,553 hours over 134,806 episodes across 195 tasks — a mean of about 95 seconds. The mean hides the important part. Individual episodes run from about 7 seconds for "put the screwdriver in the bin" to 469 seconds for "folding t-shirt pile and stacking", a sixty-seven-fold spread inside one dataset. An episode there is a whole task attempt of whatever length the task took.

In YUBI, released the same month, one episode is a recording made by a person holding a gripper with no robot in the room, and the corpus reports 8,434 hours over 1.20 million episodes — about 25 seconds each. Those episodes contain no robot joint angles at all, because there was no robot; they contain a six-degree-of-freedom end-effector pose recovered from tracking, plus a gripper width.

In AIRoA MoMa 5k, released in July 2026, one episode is not a task at all. It is a single primitive action, a sub-step of a task, kept independently addressable with metadata pointing back at the task it came from. The corpus reports 1,184,259 episodes over 5,025 hours, which is about 15 seconds each. Comparing that episode count with ABC-130K's is a six-fold unit error before any other consideration.

So one figure counts robot joint trajectories, one counts whole task attempts, one counts human hand trajectories no robot has ever executed, and one counts fragments of tasks. A "1.2 million episodes" headline conceals all of that. Whenever this document quotes a count, it says what the unit contained.

The other trap is that hours and episodes measure different things and are often quoted interchangeably. Hours measure how long somebody was working. Episodes measure how many attempts exist. A corpus of long episodes has fewer resets, and resets are where the diversity comes from, so hours flatter a dataset whose episodes are long. Where both are published, this document gives both.

2.2 The five maturity labels#

Every item below carries one of five labels. They are in descending order of how much you can actually do with the thing today.

The table reads as a promise about what you can do this afternoon, not as a quality judgement. A paper-only result can be more important than a shipped product.

A thing can carry two labels for two of its parts, and several do. The most common pattern in 2026 is a dataset that is downloadable while the rig that produced it is proprietary.

3. Teleoperation rigs, and the collapse in their price#

3.1 ALOHA, and what it made ordinary#

Before 2023, collecting fine bimanual manipulation data meant high-end arms, accurate sensors and careful calibration, which in practice meant a funded laboratory.

ALOHA changed the shape of the problem. Its claim was that a low-cost puppeteering rig, paired with a policy called Action Chunking with Transformers, learned six genuinely fine tasks — opening a translucent condiment cup, slotting a battery — at 80 to 90 per cent success from about ten minutes of demonstrations each.

The mechanism is direct joint-to-joint puppeteering. Two full-size follower arms are driven by two smaller leader arms that the operator holds, one in each hand, with four cameras recording. The operator's hands are doing the task; the robot copies. That is why a person can control fourteen joints at once, which no gamepad or keyboard scheme achieves.

It matters because it set the default recipe that everything since has reacted to, and because the data it produced is what most people's first policy is trained on.

What it cannot do is move, feel or scale. The original is bolted to a table, has no force feedback, and the ten-minute figure is for six narrow tasks in the room where the data was collected. Mobile ALOHA added a wheeled base with the operator physically tethered to it, and reported that co-training on the existing stationary corpus raised mobile task success by up to 90 per cent from 50 demonstrations per task — a best case across tasks rather than an average. ALOHA 2 redesigned the gripper and frame for durability and published the designs plus a MuJoCo model with system identification, which is the part most people now use. That tech report gives no reliability numbers of its own. The original ALOHA's dollar cost is widely quoted at around twenty thousand and does not appear in the paper; this document will not repeat a figure it could not source.

3.2 The hundred-dollar arm#

Before the community hardware line, "low cost" in this field meant five figures.

What changed is SO-101, designed by The Robot Studio with Hugging Face, whose own bill of materials prices a complete leader-and-follower pair at $229.88 in the United States or €226.30 in Europe, excluding 3D printing. A follower arm on its own is $121.94. Individual servos are $13.89.

The mechanism is the same leader-follower puppeteering ALOHA used, built from Feetech STS3215 servos instead of industrial hardware. The interesting detail is the gearing. The follower uses 1/345 reduction throughout. The leader mixes ratios — 1/191 on shoulder pan and elbow, 1/345 on shoulder lift, 1/147 on the wrist and gripper — so that the leader holds its own weight against gravity while still being easy to push around by hand. That single choice is why the rig is usable for an hour rather than a few minutes.

It matters because the price of the rig stopped being the reason not to collect data. A person can own a two-armed demonstration setup for the price of a games console, and the data lands directly in the format section 7 describes.

Three things it cannot do. The structure is 3D-printed plastic, so the payload is around a kilogram and a half at best. There is no force feedback and no torque sensing, so the leader arm is position-only and the operator feels nothing of the contact. And availability is the real constraint rather than price: of the two vendors whose prices I could read on 23 September 2026 — Seeed Studio at $200 and WowRobo at $199 — both listed every variant as out of stock. Its maturity is Downloadable under Apache-2.0 for the designs, and Shipped through third-party kit vendors when they have stock.

Two relatives are worth a line. SO-100 is the previous version and its own repository marks it deprecated, though a large installed base and many published datasets still use it. LeKiwi puts an SO-arm on a three-wheel omnidirectional base with a Raspberry Pi 5, priced in its own bill of materials at $482 for the complete 12-volt robot; its stated limitation is that keyboard base control needs a global key backend and therefore does not work on Wayland or on a headless Linux machine. XLeRobot combines two SO-101 arms with a LeKiwi base for a stated $660 total.

3.3 What a full rig costs today#

Prices in this field are quoted loosely, so the table gives only figures printed on the vendor's own product page on 23 September 2026. Read it as a ladder: each step up buys payload, reach and somebody else's assembly time.

Two notes on that table, because both are easy to get wrong.

Trossen retired the ALOHA brand in March 2025 and the legacy ALOHA line reached end of life on 1 July 2025, so the discontinued product pages carry no price. The current products still contain "Aloha" inside their full names, which causes persistent confusion. Their own pages also contradict each other on whether the WidowX AI arm has six or seven degrees of freedom, and this document could not resolve which is right.

Reachy Mini is in the table because it is the highest-volume open robot in the list and people assume it collects manipulation data. It does not. It is an expressive desktop robot with a six-degree-of-freedom head and no arms.

3.4 The three 2026 reactions#

The cheap rig succeeded, and 2026 produced three distinct responses to what it could not do.

The first says the cheap rig is too weak. MEVION, July 2026, states the complaint plainly: ALOHA-style hardware is the de facto standard but cannot generate high forces or speeds, so heavy objects and fast manipulation are out of reach. It answers with four six-degree-of-freedom arms at 7.0 kg payload and 60 N·m maximum torque, buildable for about $14,000 from parts orderable online, using sheet-metal welding for the large structure and a closed-link elbow borrowed from quadruped design to keep mass off the far end of the arm. All hardware and software are released. Its maturity is Downloadable. What it does not report is a comparison of data quality against the rig it criticises.

The second says the cheap rig should be run as a fleet. ArmnetBench, July 2026, ran an "arm farm" of SO-101 cells under light on-site supervision and compared seven policies across twelve tasks, each trained on 50 demonstrations, producing 2,518 policy rollouts and 600 reference demonstrations, every episode labelled three ways as successful, suboptimal or failed and released in LeRobot format. It matters because it is the first evidence of the $230 arm being used as evaluation infrastructure rather than as a hobby project, and evaluation is the expensive half of this discipline.

The third says the controller should be a phone. Phone2Act, May 2026, turns a commodity smartphone into a six-degree-of-freedom controller through Google ARCore and exports natively into the LeRobot format; fine-tuning GR00T-N1.5 on 130 episodes reached 90 per cent on a real multi-stage pick-and-place. COBALT, the same month, runs teleoperation in the cloud with dozens of concurrent users, sub-100-millisecond latency for up to eight users per graphics card, and a pilot dataset of over 7,500 demonstrations and 50 hours collected with smartphones across nine countries in five days. Its user study reports that phone teleoperation performed comparably to or better than specialised hardware, which is the least expected result in this document. Both are Paper only as regards a product.

Virtual-reality teleoperation sits alongside these rather than replacing them. Open-TeleVision is the reference open stack and introduced active stereoscopic feedback, where the operator's head motion moves the robot's camera so they can look around rather than watch a fixed view. The 2026 successors are honest about its ceiling. Dexora, May 2026, uses a headset for fingers only and an exoskeleton backpack for the arms, because head-mounted hand tracking is not accurate enough for arm pose — which is the most useful single sentence anyone has published about Vision Pro teleoperation. LeRobot ships supported plugins for WebXR phones and headsets, Meta Quest and PICO 4 including controller-free hand tracking, so the software is Shipped; no Vision-Pro-specific teleoperation product is for sale.

3.5 What a teleoperation rig still cannot do#

Four things, and they are why sections 4, 5 and 8 exist at all.

It needs the robot present, so data can only be collected where a robot is, at the rate one robot can be driven, by a person who is doing nothing else.

It gives the operator no sense of touch, on almost every affordable rig. Force feedback exists — the I2RT YAM active leader arm and the OpenArm line are bilateral, and the LeRobot plugin catalogue now lists the Haply Inverse3 and the Force Dimension omega.7 — but it is not what a $230 pair of arms does, and the tasks that most need it are the contact-rich ones.

It tires the operator. Mobile ALOHA's whole-body teleoperation is physically demanding, and that is a hard limit on how many hours a person can produce.

It records one embodiment. Everything collected on a particular pair of arms is tied to those arms' kinematics, and section 9.2's finding that single-embodiment data can transfer better than pooled multi-embodiment data is the one piece of good news on that front.

4. Handheld grippers: collecting without a robot#

This is the largest single change in how manipulation data is collected, and almost all of it has happened since early 2024. The idea is simple. Instead of putting a person behind a robot, put the robot's gripper in the person's hand, record what the hand does, and use that as the action label.

4.1 The Universal Manipulation Interface, and what it replaced#

Before UMI, there were two ways to get manipulation data and both had a hard ceiling. Teleoperation needed the robot present during collection, so data could only be gathered in rooms that contained a robot, at the rate one robot could be driven. Video of people doing tasks scaled without limit but carried no action labels and showed a hand rather than a gripper. Earlier instrumented handheld grippers existed, and the UMI paper is direct about why they did not take over: their recorded actions were confined to simple grasping and quasi-static pick-and-place, because end-effector tracking was not good enough for anything faster.

What changed is that demonstrations recorded by a person holding a gripper, with no robot in the room, became good enough to train visuomotor policies that deploy zero-shot on several different robots, including on dynamic, bimanual and precise tasks.

The mechanism is a 3D-printed parallel-jaw gripper with soft fingers and a trigger, carrying a GoPro as its only sensor. A 155-degree fisheye lens gives the wide view a wrist camera needs, and the paper deliberately does not rectify it, because rectifying a 155-degree view to a pinhole model stretches the periphery and compresses the centre where the useful information is. Two side mirrors sit in the camera's peripheral view and act as a pair of virtual cameras, giving implicit stereo without a second camera's cost or weight. The GoPro writes its accelerometer and gyroscope into the video file, and an inertial-monocular system built on ORB-SLAM3 recovers the six-degree-of-freedom trajectory at absolute scale. The gripper weighs 780 g and has an 80 mm finger stroke. The published cost is two numbers, not one: the printed gripper has a bill of materials of $73, and the GoPro with accessories comes to $298.

This matters because it decoupled collecting data from owning a robot, and because it standardised on one video file per episode, which is what made sharing data between laboratories plausible at all. Every item in the rest of this section descends from it.

It still cannot do three things, and the paper says all three. The kinematic limits of the robot that will eventually run the policy are unknown at collection time, so infeasible trajectories have to be filtered out afterwards. The tracking inherits visual SLAM's need for texture and fails in plain rooms with blank walls; the data scaling laws study, which collected with UMI, lost roughly ten per cent of its data this way. And collecting with the gripper is slower than using your own hands, because it is bulky and has fewer degrees of freedom. An independent user study in March 2026 (arXiv 2603.17189) put eight participants through opening sterile medical packaging with two gripper designs and with bare hands, and found both grippers substantially slower and less effective than hands.

Its maturity is Downloadable: the code and hardware files are MIT-licensed, and you print and assemble the device yourself.

4.2 The 2026 variants, and what each one added#

Through 2025 and 2026 the original design sprouted a family. Each member fixes one named weakness of the original, which makes the family a useful map of what the weaknesses actually were.

The table lists the ones with a released artefact or a measured result. Read the middle column as "the UMI limitation this attacks", and the last as what you can do with it today.

Two of these deserve more than a row.

YUBI, published in June 2026 by Toyota Motor Corporation's Frontier Research Center with AIRoA, is the first time the family acquired a reference device backed by a manufacturer. Before it, the pistol grip was the default and its ergonomic cost was anecdotal. What changed is a finger-aligned gripper whose jaw is driven directly by the operator's finger rather than by a trigger, tracked by a virtual-reality system rather than by SLAM, and a corpus behind it of 8,434 hours over 1.20 million episodes and 119 tasks. The mechanism of the corpus is worth stating precisely, because it is the part nobody can replicate casually: 179 operators working 22 desks around the clock for over two months. A single policy trained on it transferred to UR, Franka and ELEY bimanual robots by bolting the gripper onto each. It matters because the hardware licence is unusually permissive for a manufacturer — the hardware is CERN-OHL-W-2.0 with an explicit note permitting commercial manufacture and sale, and the software is Apache-2.0. What it cannot do yet is give you the data: the full corpus was still marked as coming soon in September 2026, and only a gated subset exists.

HiFi-UMI, published in July 2026 by Simple AI, is the most consequential result in this whole section. Before it, standard practice treated robot-free handheld data as suitable for pretraining only, with a small quantity of real teleoperated data added at the end to anchor the policy to the robot — because the handheld trajectories were not accurate enough to train on alone. What changed is the claim that raising the fidelity of the handheld data removes the need for that anchor entirely. A policy post-trained only on handheld demonstrations deployed directly on a real robot and matched in-domain teleoperation across three different model backbones, with success-rate differences of −2.5, +3.1 and −0.6 percentage points; the strongest reached 85 % on a precision insertion task, and the teleoperation baseline it was compared against had been collected in the evaluation scene while no handheld trajectory had. The mechanism is four hardware decisions: head-mounted offline stereo-inertial SLAM rather than wrist SLAM, natively measured rather than reconstructed pose between the two grippers, a shared microsecond-level trigger synchronising the sensors, and two wide-angle cameras per hand covering about 200 degrees. Together they reach 3 mm workspace-local end-effector accuracy with no external tracking hardware in the room.

One honest qualification belongs with that result, and the paper makes it. The headline comparison used 3,200 handheld trajectories per task against 300 teleoperated ones. It is a comparison of two data pipelines at their natural throughputs, not a claim that one handheld demonstration is worth as much as one teleoperated demonstration. Comparing pipelines is the right comparison to make, because the whole argument is that one pipeline is far cheaper per hour than the other, but it is not the comparison people will remember.

It matters because it removes the last structural reason to own a robot in order to produce a deployable policy for a task. What it cannot do is let you build the rig: the hardware is proprietary, and what you can have is 2,000 hours of its data under CC BY 4.0. The accuracy figure is also workspace-local rather than global, which is the right number for a tabletop task and not for a mobile one. Its maturity is therefore split: the dataset is Downloadable, the rig is Demonstrated.

4.3 The scale nobody had claimed before#

One July 2026 result is worth recording separately, because of what it implies about where this is going.

Before it, the largest published handheld corpora were measured in thousands of hours. Xiaomi-Robotics-1 reports pretraining a vision-language-action policy on over 100,000 hours of real-world manipulation trajectories collected with UMI devices. The mechanism that makes that quantity usable is an automatic labelling pipeline that annotates trajectory clips with sentences describing how the scene changed, so that the actions have language to condition on without anybody writing it. The paper reports that performance improves consistently with both data scale and model size during pretraining, and that the improvement carries through to out-of-the-box performance on real robots in unseen rooms.

It matters because it is the first evidence that handheld collection scales into the range where the word "foundation model" is not aspirational, and because that scale was reached without a robot fleet.

What it cannot do is let anybody check it. Neither the corpus nor the collection infrastructure is released, the comparison is against the authors' own baselines plus simulation benchmarks, and 100,000 hours is not a quantity a reader can sanity-check. Its maturity is Paper only.

4.4 What you can buy, and what it costs#

Until 2026 a handheld rig meant a printer and a weekend. Two vendors now sell assembled ones, and the prices are worth setting against UMI's own $73 plus $298.

The table lists only devices with a published price and stock on the vendor's own page as of 23 September 2026. Read the price column as what leaves the shop, and the last column as what you still have to add.

TRumi replaces UMI's printed gear trigger with a zero-backlash cam drive on linear rails with constant-force springs and a 0.4 lb pull, and widens the field of view from 155 to 177 degrees. Mantis matches its gripper geometry to the vendor's own arm so that data transfers directly. Grabette, announced in July 2026, is not a product at all but a published design — two cameras including a depth camera, an inertial sensor, magnetic encoders and a Raspberry Pi — that writes straight into the LeRobot dataset format as a camera-local six-degree-of-freedom pose plus gripper state.

The honest summary is that the printed original is still an order of magnitude cheaper than anything assembled, and what the premium buys is backlash-free mechanics, a wider lens and somebody else's assembly time.

One device that is frequently written about is not buyable. BeingBeyond announced U1, described as the first commercial dexterous-hand handheld rig, in March 2026, with claimed figures of 680 g, eleven joints of which six are active, 5 kg payload, sub-millimetre tracking error and tactile sensing on all five fingertips. Its research basis, RealDexUMI, reports 88.75 % average success across eight real-robot tasks. There is no published price and no purchase path; the product page offers only a consultation booking. Its maturity is Announced, and the stated limitation from its own project page is that the in-hand camera view limits global scene awareness, so it fails when the target or a progress cue falls outside that view.

4.5 The shared evaluation floor that did not exist#

One thing the handheld family conspicuously lacked until this autumn was any way to compare two claimed success rates. Every group evaluated on its own robot, objects and tasks.

The UMI Arena, a CoRL 2026 workshop scheduled for 12 November 2026 and run by AIRoA, is the first attempt at a common floor. It runs three tracks: a policy competition judged by the organisers on real Franka, TX-G2 or OpenArm arms, a device exhibition judged by attendee vote on ergonomics, dexterity, strength, reproducibility and cost, and a paper track. The competition uses eight tasks of which three are held out until final evaluation, so the leaderboard measures generalisation rather than practice. Submissions opened on 21 September 2026 and registration closes on 9 October.

It matters because a claimed success rate in this family has until now been uninterpretable. What it cannot do is serve as a neutral standard: the evaluation runs on the organisers' robots, against a dataset collected with organiser-affiliated hardware, and that dataset is gated behind a signed agreement whose terms are themselves behind the gate. Three of the eight tasks are secret. Its maturity is Announced, and no results exist yet.

5. Exoskeletons, gloves and ambient capture#

A handheld gripper records where the gripper went. It records nothing about the arm that carried it, and nothing about the fingers. This section is the three ways people record those instead: a frame worn on the arm, a glove worn on the hand, and a room instrumented to watch the whole body.

5.1 Exoskeletons, which record the whole arm#

Before exoskeletons, the choice was teleoperation, which is expensive, or a handheld gripper, which captures the end-effector only. Behaviours that use the whole arm — gathering objects towards you, supporting something with your forearm — cannot be recorded by a device that only knows where your hand is.

AirExo, 2023, is the origin of this line. Its claim was that three minutes of teleoperated demonstrations, augmented with exoskeleton data collected outside the laboratory, matched or beat a policy trained on over twenty minutes of teleoperation, and was more robust to disturbance. The mechanism is a passive dual-arm frame worn by the person, carrying joint encoders and nothing else, at a stated $300 per arm. Its code and CAD are MIT-licensed, and its maturity is Downloadable. Its limits are that the comparison is on a small task set, that a passive frame gives the wearer no force feedback, and that it needs adapting to each robot's kinematics.

AirExo-2, 2025, made the cost argument explicit: $600 for the exoskeleton system against roughly $60,000 for a teleoperation platform, a hundredfold ratio. Its contribution beyond price is the idea of an adaptor, a preprocessing step that converts captured human motion into pseudo-robot demonstrations, so no robot fine-tuning stage is needed afterwards. A policy trained solely on adapted in-the-wild data reached performance comparable to one trained on teleoperation — against the authors' own baseline on their own tasks. Its maturity is Downloadable, with a caveat that matters: the repository's licence resolves to "no assertion", so the CAD terms are unclear.

Three 2026 papers push the line further and each attacks a different weakness. ExoGS, January 2026, introduces AirExo-3, a robot-isomorphic passive exoskeleton, and rebuilds the robot, objects and environment as editable Gaussian-splatting assets so one human demonstration can be replayed and augmented in simulation; its repository carries no licence file at all. The Universal Manipulation Exoskeleton, June 2026, adds real-time haptic torque feedback and records joint torques as well as configurations, with the striking demonstration that operators could unsheathe kinematically constrained objects while blindfolded; it publishes no cost, no degree-of-freedom count and no success rates, so its maturity is Demonstrated. DexEXO, March 2026, attacks fit, supporting hand lengths from 140 mm to 217 mm so that many different people can use one rig, and gives the exoskeleton a passive hand that visually matches the deployed robot, which removes the image-inpainting step that DexUMI needs.

For humanoids specifically, HOMIE is the one to know. It is an isomorphic exoskeleton cockpit at a stated $500, with a reinforcement-learning locomotion policy driven by a foot pedal, isomorphic exoskeleton arms, and gloves that use Hall sensors rather than servos so that a compact glove reaches more than fifteen degrees of freedom. It claims half the task completion time of prior teleoperation systems. What it cannot give you is direct control of the legs: the body is policy-driven, so the operator commands a gait rather than a stride.

5.2 Gloves, which record the fingers#

Before 2024, hand motion capture was not portable and did not turn into policies. DexCap changed that by combining SLAM for the wrist with electromagnetic sensing for the fingers, which is the part that survives occlusion, and publishing an imitation algorithm that consumed the result. Its weakness is inherent to the sensing: electromagnetic tracking degrades near metal, and retargeting to a non-human hand is still required afterwards.

The 2025 and 2026 glove work converges on one idea, and naming it is more useful than listing the devices. The embodiment gap is being closed in hardware rather than in software. Rather than recording a human hand and then correcting for the fact that the robot's hand is different, these systems arrange for the two to be the same thing, or to look the same, or to show the operator what the robot would have felt.

The table lists the open ones with a published number. Read the middle column as the specific gap each closes.

The commercial side is thinner than it looks. MANUS is the clearest vendor positioning itself at robot training, with Metagloves Pro and Metagloves Pro Haptic listed at €4,500 and the older Quantum Metagloves at €2,500 and marked end of life; those figures came from the vendor's landing page rather than a checkout page. A new MANUS product is announced for IROS 2026, which opens four days after this document's date, so anyone choosing a commercial glove should wait a week. Rokoko's glove product page returns a 404 and this document could not establish what, if anything, replaced it.

5.3 Ambient capture, which records the room#

The newest idea in this section is to stop instrumenting the person and instrument the building.

Before it, every capture method fragmented the experience. One rig recorded the hand, another the body, another the objects, and nothing recorded how they evolved together while a person pursued a goal over several minutes.

ACE-Data-0, July 2026, turns real homes into calibrated, synchronised recording studios at two scales: a table-scale configuration that resolves hand-object manipulation and a room-scale configuration that captures walking, whole-body motion and interaction across a furnished home. It records egocentric and multi-view exocentric video, full-body and articulated hand motion, object geometry and six-degree-of-freedom object trajectories, audio and touch as one stream. The planned release is 150 hours, 17 million frames, 75,000 interaction episodes over 200 task categories, performed by 50 participants in two environments, with individual takes running up to 20 to 30 minutes. Participants are given goal-level instructions rather than step-by-step ones, deliberately, to preserve natural variation.

It matters because it is the only method here that captures the whole perception-action loop of a person doing a real task in a real house, which is the thing every other method approximates.

Two things it cannot do. The data does not exist yet: the dataset card says the figures describe the planned release and the files are marked as coming soon, so its maturity is Announced. And when it arrives it will be research-only — the licence is a bespoke one granting non-commercial academic use, with manual per-person approval and no redistribution. Do not plan a product around it.

6. The open datasets, and their licences#

6.1 Why the licence comes before the size#

In most areas of this repository a licence is a thing you check before shipping. Here it is the first thing to check, before you download anything, and there are two reasons.

The first is that a dataset licence propagates into the weights. A model trained on a corpus published under a non-commercial licence is a derivative of that corpus, and several of the largest and most attractive robot datasets are published exactly that way. Discovering this after a training run has cost you the training run.

The second is that there is no index. I looked for a current, maintained register of robot manipulation datasets with their sizes and licences, and there is not one. Open X-Embodiment's spreadsheet is the closest thing, and it has no licence column at all. The two community lists on GitHub have one star and zero stars respectively. Anyone writing about this has to read the dataset cards one at a time, and so does anyone using them.

6.2 Open X-Embodiment, the pool that stopped growing#

Before it, every laboratory's dataset was an island. There was no way to train one model on data from several robots because there was no common format and no common place to put it.

What changed in October 2023 is that 34 institutions agreed to convert their data into one format and publish it together. The current index lists 72 member datasets, 2,419,193 episodes, 27 distinct robots and about 8,965 GB.

The mechanism was conversion rather than collection, and that is where the interesting detail sits. The index records how each member dataset was gathered, and the answer is not what the phrase "robot demonstrations" suggests. Twenty-five member datasets were collected by a person in virtual reality, fifteen by a script, thirteen by an expert policy, six with a space mouse, four by puppeteering and two by moving the arm by hand. The two largest members are not human demonstrations at all: VIMA contributes 660,103 scripted episodes and QT-Opt contributes 580,392 episodes generated by a learned policy, and between them those two are about half of every episode in the pool.

It matters because it is the reason cross-embodiment models exist at all, and because it proved that laboratories would pool data if somebody did the conversion work.

Three things it cannot do. There is no global definition of an episode, so the unit varies across members from a three-hertz scripted rollout to a thirty-hertz human demonstration. There is no central licence record, so a commercial user has to clear 72 licences individually, and at least one member — the non-commercial half of RH20T — is CC BY-NC 4.0 and sits inside mixtures people train on without noticing. And it has effectively stopped: the maintaining repository was last updated in November 2025. Its maturity is Downloadable, with the licence unresolvable centrally.

6.3 DROID, and a licence that is simply missing#

Before DROID, large manipulation datasets came from one laboratory's robot in one laboratory's building.

DROID, published in 2024, collected 76,000 teleoperated trajectories totalling 350 hours across 564 scenes and 84 tasks, using 50 data collectors across North America, Asia and Europe over twelve months. The mechanism was a standardised mobile cart carrying a Franka Panda, so that the same rig could be wheeled into a genuinely different room. One episode is one teleoperated trajectory with up to three alternative language instructions and a seven-dimensional action.

It matters because scene diversity, not episode count, is what section 9 says buys generalisation, and DROID was the first serious attempt to buy scene diversity deliberately.

What it cannot give you is a licence. There is no licence file in the dataset repository, no licence statement in its documentation, and none in the storage bucket; only the separate policy-learning code is MIT. A commercial user cannot clear DROID from public documents, and this document is not going to guess on their behalf. Its documentation also records that about twenty per cent of episodes were lost while face-blurring and copying the raw release, which affects the raw video version and not the converted one. No successor had appeared by September 2026. Its maturity is Downloadable, licence unnamed.

6.4 The 2026 wave#

Six large corpora arrived between June and September 2026, and they changed what is available far more than anything in the two preceding years. The table lists them with the three numbers that matter and the licence read from the dataset card. Read the "one episode is" column first, because it is what makes the size column mean anything.

Two of them deserve the full treatment.

ABC-130K is the most useful thing released this year for somebody who wants to train on open data and sell the result. Before it, DROID's 350 hours of single-arm Franka data was the reference open teleoperation set and nothing open was bimanual at scale. What changed is a complete open stack: 3,553 hours of two-armed teleoperation, plus 400 hours of simulated teleoperation across 20 scenes, plus over 100 hours of real evaluation runs published with their scoring rubrics, plus the hardware design, the training code and the simulation pipeline. The mechanism is a standardised station: two I2RT YAM arms on an aluminium extrusion rail, three Intel RealSense D405 cameras — one on a metre-high post and one on each wrist — inside a cage whose dimensions replicate the simulated scene exactly, so that a simulated evaluation predicts the real one. The 195 tasks fall into seven categories: pick-and-place, folding, sorting, tool use, handover, insertion and assembly. It matters because the correlation between the simulated and real evaluations gives you a way to compare model choices without running a real robot each time, which is the expensive part of this whole discipline. What it cannot tell you is what the rig costs: the hardware page is a full bill of materials with live supplier links and no total anywhere, and neither the paper nor the site states one. Its maturity is Downloadable under Apache-2.0, with a click-through gate that adds no terms.

HIW-500 is the answer to a different question. Before it, humanoid data was collected in laboratories and staged environments. What changed is 500-plus hours of whole-body teleoperation of a Unitree G1 recorded in twelve real homes in Southeast Asia, where the layout, the object states, the lighting, the clutter and the operator's style all vary between episodes. The mechanism is ordinary whole-body teleoperation with head and wrist cameras at 480p and 30 frames a second, 29 joint states, inertial and odometry data, and 148,000 subtask annotations drawn from a vocabulary of 161 labels. It matters because it is the largest humanoid corpus a commercial user may actually use, released under CC BY 4.0 with attribution and no gate. What it cannot offer is task breadth: the diversity is in the twelve homes, and the card claims only "10+ tasks".

Three more are worth naming without the full treatment. AgiBot World remains the largest single-programme collection, with 1,003,672 trajectories in its Beta release across 217 tasks and 100 robots, and it is published under CC BY-NC-SA 4.0, which puts the largest robot dataset in the world out of reach of a product. RoboMIND has 55,000 teleoperated trajectories over 279 tasks and four embodiments under Apache-2.0, which makes it unusually well-placed for its size. And the Galaxea Open-World Dataset has 500-plus hours of mobile manipulation on a single uniform embodiment, under CC BY-NC-SA 4.0.

6.5 Which licences stop a commercial user#

This is the section to read twice. Every entry was read from the dataset card or the licence file. The table sorts by whether you can ship a model trained on the data. Read the middle column as the exact instrument, not a paraphrase.

Three practical warnings follow from that table.

The non-commercial licences are concentrated at the top of the size rankings. AgiBot World, Galaxea and OpenNeoData are among the largest things available and none of them may be used in a product. If your plan is "train on the biggest open dataset", your plan currently ends in a licence review.

ShareAlike is not the same as non-commercial and is often worse in practice. CC BY-SA and CC BY-NC-SA both carry an obligation to license derivatives on the same terms, and a model is a derivative. A company that can live with attribution frequently cannot live with publishing its weights.

A mixture is only as permissive as its most restrictive member. Training on a blend that includes Open X-Embodiment means inheriting whatever is in it, including the non-commercial half of RH20T, and there is no central record to check against.

6.6 One headline that does not survive arithmetic#

Unitree's whole-body teleoperation release has been described as the largest open humanoid teleoperation dataset. It is worth working through, because the shape of the error is common.

There is no single repository of that name. What exists is a collection pointing at 59 separate per-task repositories on the Hugging Face Hub. Summing the frame counts in each repository's own metadata file, at the 30 frames a second those files declare, the whole-body teleoperation group comes to roughly 12,900 episodes and about 136 hours. Unitree's entire public dataset corpus, including every other robot and end effector, comes to about 496 hours.

HIW-500, on the same robot, is 500-plus hours. So the whole-body release is around a quarter the size of a single other humanoid dataset, and Unitree's complete public output is smaller than that one dataset. The claim is only defensible if it is read as counting tasks rather than hours, or if it counts data that has not been released.

The lesson generalises. A dataset distributed as many repositories has no single number attached to it, which means the number in the announcement came from somewhere other than the files. The check takes an hour and is worth doing before you plan around a figure. The Unitree data itself is genuinely useful and genuinely permissive — Apache-2.0, ungated — and none of that is in question.

7. LeRobot, and the tooling that grew round it#

7.1 What it is, and why it won#

Before LeRobot, every laboratory that collected demonstrations invented its own file layout. A dataset from one group could not be loaded by another group's training code without a conversion script, and the conversion script was usually the thing that got lost. Open X-Embodiment's answer in 2023 was to convert sixty datasets into a single TensorFlow-based format, which worked but tied everything to one framework and one storage style.

What changed is that LeRobot, Hugging Face's robot-learning library, became the place where the format, the training code, the pretrained policies, the hardware drivers and the published datasets all live together. As of 23 September 2026 the repository has about 27,700 stars and is under Apache-2.0, which is the permissive end of the spectrum and one reason it spread.

The mechanism is the LeRobotDataset format, which stores frames in Parquet files, pictures go into MP4 shards, and a meta/ folder carries the frame rate, the field shapes, the normalisation statistics and the task sentences. Because pictures are stored as video, a few hundred episodes are megabytes rather than gigabytes, and because everything sits on the Hugging Face Hub, publishing a dataset is a push rather than a hosting project.

This matters more than a file format usually would. The entire community hardware line in section 3 is only useful because the data it produces lands somewhere that training code already reads.

What it cannot do is make datasets comparable. Two LeRobot datasets can share every key name and still differ in frame rate, camera placement, action convention and what counts as an episode boundary, and nothing in the format records the difference.

7.2 What the 2026 releases added#

The library shipped four releases in 2026 that are worth knowing about, and two of them landed after May.

The table gives the dates from the repository's own release list and the change that matters most to someone collecting data. Read it as a compatibility warning as much as a feature list, because each of these releases broke something.

The v0.6.0 release is the substantial one, and its release notes are worth reading in full before upgrading. Three changes matter to anyone collecting.

Depth is now a first-class citizen rather than something you bolted on, with Intel RealSense integration and separate codec settings for colour and depth streams. Before this, most people either discarded depth or stored it as a side-car that the training code did not know about.

Episodes can now be annotated with language automatically. The lerobot-annotate command runs a vision-language model over the recorded video and writes the task sentences, replacing the per-frame subtask_index annotations that earlier versions used. This matters because hand-writing a sentence per episode is one of the genuinely tedious parts of preparing a dataset, and because vision-language-action policies need those sentences.

Data loading is about twice as fast, through decoding several camera streams in parallel. Video decoding was already the slow part of training a policy on image data, so this is a real change to how long an experiment takes rather than a benchmark improvement.

Two things v0.6.0 will break. Installing lerobot no longer pulls in dataset or training dependencies, so pip install lerobot[training] is now required; the base install dropped roughly forty per cent of its dependencies. And the minimum PyTorch version moved to 2.7, which on an older machine is not a small ask.

Its maturity is Shipped and Downloadable under Apache-2.0.

7.3 The size of the pile#

One artefact gives the scale of what the format unlocked. In July 2026 the LeRobot project published community_dataset_v3, an aggregate of 791 datasets covering 46 robot types from 235 community contributors, under Apache-2.0 and with no gate. Each of those datasets is small. Most were recorded on one cheap arm by one person in one room, and most will never be used by anybody but their author.

That is not a criticism, and it is worth being clear about why. The value of the pile is not that any one dataset in it is good. It is that the format has enough gravity that publishing became the default rather than an afterthought, and section 9 is about the conditions under which many small, varied datasets are worth more than one large uniform one. On the evidence there, 791 rooms is the kind of diversity the scaling results say you want, and 235 different people holding the controls is the kind they warn you about.

8. Human video instead of teleoperation#

8.1 Why the idea keeps coming back#

The arithmetic is unavoidable. Pooled robot data across the whole field is measured in millions of episodes. Video of people manipulating objects is effectively unbounded, and a large amount of it is already recorded, already egocentric, and already public. If any usable fraction of that could substitute for teleoperation, the constraint in section 1 would stop binding.

The reason it does not simply work is that a video of a person contains no actions. It shows a hand, not a gripper, moving under a body the robot does not have, filmed from a viewpoint the robot's cameras do not occupy. Turning it into supervision means inventing the missing action labels, and the quality of that invention is the whole subject.

Through 2025 this was, in the honest assessment the repo's own mechanism document gave it, interesting rather than proven, with almost nothing released. That is the part which changed in 2026.

8.2 The result that changed the argument#

Before June 2026, the defensible position was that human video is a useful source of visual priors and a poor source of actions, so you pretrain on it and then collect robot data for everything that matters.

HumanScale, published on 18 June 2026, claims something stronger. Holding the amount of pretraining data fixed and holding the post-training and validation protocols fixed, a model pretrained on egocentric human video beat the same model pretrained on teleoperated real-robot trajectories: 24 % lower validation loss on real-robot action prediction, 52.5 % higher success on in-distribution real-robot tasks and 90 % higher on out-of-distribution ones.

The mechanism is not a new architecture. It is a filtering and labelling pipeline applied to the human video before pretraining, and the paper's own framing is that the pipeline is the contribution: egocentric data works as a pretraining source when processed carefully, and the recipe is to pretrain on human video for diverse world representations and then adapt with a small amount of labelled robot data purely for action-space alignment.

This matters because it inverts the default. If it holds, the expensive robot data is a small alignment step at the end rather than the bulk of the budget.

What it cannot do is remove robot data. The claim is explicitly about pretraining, with real-robot adaptation still required, and the comparison holds the data quantity fixed rather than the collection cost — which is the more favourable framing for human video, since human video is far cheaper per hour. Its maturity is Paper only as regards the pipeline; the claim has not yet been independently replicated at the time of writing.

8.3 The three mechanisms people actually use#

Underneath the individual papers there are only three ways to get supervision out of a video of a person, and knowing which one a paper uses tells you most of what you need.

The table names each route and what it actually supplies to the policy. Read the last column as the thing that still has to come from somewhere else.

Four 2026 releases sit on those routes, and each is worth a line with its number.

HuRo, September 2026, takes the retargeting route at scale. It converts heterogeneous human videos into robot-aligned observations and action trajectories and publishes a dataset of about 630,000 robotized episodes and 142 million processed frames drawn from five human-video sources. On four real manipulation tasks, increasing the amount of robotized video raised overall completion from 51.5 % to 80.3 %, and completion under spatial and visual shift from 34.9 % to 72.2 %. Here an "episode" is a segment of human video with inferred robot actions attached, which no robot has ever executed. Its maturity is Paper only as regards an obtainable dataset at the time of writing.

RoboEdit, August 2026, takes the editing route. It reconstructs and retargets three-dimensional interactions from ordinary RGB video and rewrites the video so the human hand becomes a robot, producing 174,000 aligned video pairs totalling 14 million frames across seven robot embodiments. An "aligned pair" here is one human clip and its edited robot counterpart, not a robot trajectory.

ACE-Ego-0, June 2026, is the most careful about mixing. It trains on 4,530 hours of robot and simulation data together with 1,480 hours of egocentric human data converted to pseudo-actions, and adds a reliability-weighted objective so that the noisy human labels contribute less where they are least trustworthy. The interesting part is the admission built into the method: the human labels are known to be wrong some of the time, and the training has to be told so.

UMI-Bridge, September 2026, uses the handheld gripper of section 4 as a translator between human video and robot data, aligning them by equivalence of action rather than similarity of pixels. It reports 91.7 % mean success against 73.3 % for naive co-training of human and robot data, and matches a robot-only baseline using a quarter of the robot demonstrations.

If you want the field's own summary rather than four results, the survey Robot Learning from Human Videos, April 2026, is the current one.

8.4 What human video still cannot supply#

Three things, and they are the same three every paper in the area runs into.

Force is absent. A video shows where a hand went and not how hard it pressed, and for the tasks in the gripping area that are about contact rather than position, that is the signal you needed.

The hand is not a gripper. A person manipulating an object uses fingers, wrist rotation and the other hand in ways a parallel jaw cannot reproduce, so a large fraction of any human video corpus shows behaviour that is not executable and has to be filtered out or silently mislearned.

Reachability is unverified. A retargeted trajectory can be geometrically sensible and kinematically impossible for the robot that will run it, and nothing in the video says so. This is the same limitation the handheld grippers have, and for the same reason: the robot was not there.

9. Data scaling laws, and what they actually say#

9.1 The one published law#

Before October 2024, the honest answer to "how many demonstrations do I need" was that nobody knew, and the practical answer was to train on 50, then 100, then 200, and plot the curve.

Data Scaling Laws in Imitation Learning for Robotic Manipulation is still the reference study, and its last revision is dated 26 June 2026. It collected over 40,000 demonstrations and ran more than 15,000 real-world robot rollouts under a fixed evaluation protocol, varying three things independently: the number of training environments, the number of objects, and the number of demonstrations per environment-object pair.

Its central claim is that generalisation follows a power law in the diversity of the data and saturates in the quantity. The fitted relationships, where the optimality gap is the distance from perfect performance, are that the gap falls as the number of objects to the power −0.32, as the number of environments to the power −0.26, and as the number of environment-object pairs to the power −0.35. Against that, performance plateaus once the number of demonstrations reaches about 800 in their largest setting, and the recommendation is 50 demonstrations per environment-object pair and no more.

The recipe that follows is the memorable part. Thirty-two environment-object pairs, one distinct object per environment, 50 demonstrations each — about 1,600 demonstrations in total — collected by four people in one afternoon, reached roughly 90 % success on novel environments with unseen objects. The data was collected with UMI handheld grippers and the policy trained was Diffusion Policy with a DINOv2 ViT-L/14 visual encoder. One demonstration there is one short handheld recording of one task in one room with one object, so the numbers transfer directly to anybody collecting the same way and need adjusting for anybody whose episodes are longer.

This matters because it converts a vague instinct into a budget. If you are collecting in one room with three objects, another hundred episodes is close to worthless and a fourth room is worth a great deal.

The paper's own stated limitations are unusually clear, and they bound the claim tightly. It studies single-task policies, not generalisation across tasks. It studies imitation learning only, not reinforcement learning. It uses UMI's demonstrations including their errors and does not study the effect of data quality. It tests one policy algorithm. And it validates on four tasks, which the authors attribute to resource limits. Its maturity is Paper only as a law, though the code is MIT-licensed and downloadable.

9.2 The results that complicate it#

Two later studies pull in different directions, and both are worth knowing because they are the strongest available check on the headline.

Is Diversity All You Need for Scalable Robotic Manipulation?, July 2025, takes the diversity conclusion apart along three axes. It finds that task diversity does matter more than demonstrations per task, which agrees. It finds that multi-embodiment pretraining data is optional, and that models trained on high-quality single-embodiment data transfer to other platforms with better scaling behaviour during fine-tuning — which cuts against the assumption behind pooled cross-embodiment corpora. And it finds that diversity of demonstrator is actively harmful: different people's operating preferences produce a multimodality in velocity that confuses the policy. Correcting for that gave a 15 % gain, which the authors equate to using 2.5 times as much pretraining data.

Scaling Bimanual Household Manipulation, September 2026, reports the opposite of saturation on its own corpus: releasing 1,500 hours of two-armed household demonstrations, it finds task success rising steadily with both the quantity of expert data and with added on-policy correction data, "at our current data scale". The qualifier is the honest part.

The two results are not contradictory, and the reconciliation is the useful lesson. Saturation in the 2024 study was measured for a single task in a fixed set of environments. Continued scaling in the 2026 study is measured across many tasks. Quantity saturates within a narrow distribution and diversity does not, which is the same statement twice.

9.3 What to do with this#

Three practical consequences, in the order they will save you time.

Count your environment-object pairs before you count your episodes. If that number is small, collecting more episodes is the wrong move and rearranging the room is the right one.

Fifty demonstrations per pair is a defensible default, and it is the same number the two-arm benchmark leaderboards settled on independently.

Be careful about how many different people collect. The velocity-multimodality result means that three demonstrators who each have a personal style can be worse than one, which is the opposite of the instinct that more collectors means more diversity. Diversity of scene helps; diversity of operator hurts.

10. Making demonstrations instead of collecting them#

10.1 Replaying a few demonstrations into thousands#

Before MimicGen, generating training data in simulation meant scripting a controller for each task, which is exactly the work that policy learning was supposed to remove.

MimicGen, 2023, changed that by taking a small number of human demonstrations and adapting each one to new object poses, producing over 50,000 demonstrations across 18 tasks from about 200 human demonstrations. Its two-armed successor DexMimicGen, 2024, generated 21,000 demonstrations across nine bimanual dexterous tasks from just 60 source human demonstrations.

The mechanism is segment-and-transform. A demonstration is cut into object-centric segments, each segment is rewritten relative to wherever the object now is, and the result is replayed in the simulator; a replay that fails is discarded. The demonstrations it produces are therefore not new strategies, they are the same strategy relocated.

This matters because it makes the cost of a dataset depend on the number of strategies you need rather than the number of situations you want covered, which is the right way round.

What it cannot do is invent a behaviour the seed demonstrations did not contain, and it only works where a simulator can check the replay, which excludes deformable objects, most contact-rich assembly and anything whose success depends on force. Both tools are also research-only under NVIDIA's own licence rather than a standard open-source one — the GitHub API reports the licence for both as "Other" — so read the file before building a product on either. Their maturity is Downloadable, with that restriction.

10.2 The simulation benchmarks that now ship the data#

Two large simulated corpora are worth naming because LeRobot v0.6.0 wired both into its evaluation command, which is the practical sign that they have become standard.

RoboCasa supplies kitchen scenes with thousands of 3D assets across more than 150 object categories, and its 2026 incarnation RoboCasa365 supplies 365 tasks. RoboTwin 2.0 supplies 50 bimanual tasks with over 100,000 generated trajectories and strong domain randomisation; the RoboTwin platform repository is MIT-licensed. Both are Downloadable.

The limitation both share is worth stating plainly: a generated demonstration is only as good as the program that generated it, and its contact behaviour is whatever the controller did, which is often unlike a person. For a task whose difficulty is in the last centimetre, that is exactly the part the generator gets wrong.

10.3 Two 2026 ideas that change the collection loop itself#

AXIS, July 2026, is a community data engine rather than a dataset. Before it, contributing to a shared corpus meant owning the specific hardware the corpus was defined on. AXIS runs teleoperation in a browser, generates and validates new tasks automatically, and passes community-collected demonstrations through automatic success checking, quality filtering, trajectory smoothing and visual and physical augmentation before they enter the training set. It currently holds 207 tasks and over 50,000 trajectories, and continual pretraining on it improved π0.5's overall success by 5.8 points. Its maturity is Paper only as regards an obtainable release at the time of writing.

RoboPocket, March 2026, attacks a different waste. Handheld collection is open-loop: the collector does not know where the policy is weak, so coverage of the states that matter is accidental. RoboPocket runs the current policy remotely during collection and draws its predicted trajectory into the collector's phone screen in augmented reality, so the person can see a failure coming and deliberately collect there. It reports roughly double the data efficiency of offline scaling. The same instinct appears in HIL-UMI, September 2026, which triggers collection only where an energy score says the policy is out of its depth.

Both matter for the same reason, and it is the reason section 9 gives: if quantity saturates and coverage does not, then knowing where to collect is worth more than collecting faster.

11. What changed between May and September 2026#

Four months is a short window and an unusually dense one. The table lists what landed in it, in date order, with the label that says what you can do with each. Read the last column as the reason it is on the list rather than as a summary.

Read down that list and one shape emerges. Nothing in it is a new policy architecture. Almost everything is either a way to collect without a robot, a way to make already-collected data trainable, or a corpus published under a licence that lets somebody else use it. The field spent the summer on its data supply.

The second shape is a divergence between two claims that cannot both stay true. HiFi-UMI and HumanScale, six weeks apart, argue that robot-free data is sufficient and that human video is superior. PrimeBot, in September, argues that more teleoperated robot data keeps paying. Both are honest about their evidence and both may be right about their own setting, and the resolution will be an empirical question about which tasks fall on which side of the line. It is the most interesting open question in this document.

12. What none of this fixes#

Five things survive everything above, and a development that claims to fix one of them deserves more scrutiny than it will usually get.

Nobody can compare two numbers. Section 2.1 is not a pedantic preamble, it is the state of the field. An episode count means nothing without knowing what an episode was, an hours count flatters long episodes, and there is no maintained index that records either. Until there is, every league table of dataset sizes is a comparison of incompatible units, and the UMI Arena is a workshop competition rather than a standard.

The robot was not there. Every route that removes the robot from collection — handheld grippers, exoskeletons, human video — produces trajectories that nothing has verified the robot can execute. The kinematic limits are unknown at collection time, so infeasible motions have to be filtered afterwards, and the filter is a heuristic. This is the single limitation shared by section 4, section 5 and section 8, and it is structural rather than an engineering gap.

Force is still missing from almost everything. Colour video and joint angles are cheap to record and are what nearly every corpus contains. How hard the gripper pressed is expensive to record, is absent from every large open dataset except RH20T and a handful of research rigs, and is exactly the signal the gripping area says decides whether contact-rich tasks work. Adding a force channel to a dataset format is easy; adding it to three thousand hours of recordings that were made without a force sensor is impossible.

Evaluation costs more than collection. Every claim in this document about success rates rests on real-robot rollouts, and those are slow, are run by the authors, and are almost never reproduced. ABC-130K publishing its evaluation logs and rubrics is notable precisely because it is unusual. The reason simulated evaluation gets so much attention is not that anyone thinks it is accurate; it is that the real thing does not scale.

Cheaper collection does not make collection free. The hardware price has fallen by more than an order of magnitude, and the cost that remains is a person's attention, one episode at a time. The scaling results in section 9 say what that person should spend their attention on — more rooms and more objects, not more repetitions — and that is a change in what you do, not a change in how long it takes. A week of collecting is still a week.