Executive Summary

Key numbers

Four cells. The first two cover the size of this data and the precision of the rework; the second two, its contribution to the score and how far it can be checked from outside.

Source: Lambda-humanoid Team, "λ₀: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation" (arXiv:2610.18950, 13 October 2026) and the OpenLambda repository.

To put a humanoid to work around a house, walking and grasping have to run as one motion. The standard way of gathering that kind of data has been whole-body teleoperation. A person straps on a rig and moves the robot's arms and legs in place of their own, producing one demonstration at a time. The paper writes down the limit of that approach. Because an operator has to walk a physical humanoid through every episode, the volume collected is tied to robot availability, and widening the range of tasks and environments grows expensive.

First-person video carries no such constraint. People pick things up, carry them and put them away all day long with no robot in the room. The trouble is that most existing egocentric data kept only the wrist and hand trajectory. Where the person walked, how far the knees bent, which way the torso leaned: that part is missing from the record. For a task where walking and grasping are a single motion, the missing half is the more important one.

HumanVerse is a 500-hour record built to fill that gap. It holds 32,993 episodes and 87 million pose frames. The people filming wore a GoPro HERO13 camera, a Pico 4 Ultra headset and whole-body trackers at once. Video and body motion are logged against the same clock, and finger motion is recovered from the video afterwards. A camera of the same model sits on the head of the Unitree G1, so the field of view a person saw lines up with the one the robot will see.

The filming happened in homes, retail shops and supermarkets, industrial sites, offices and outdoors. The work is ordinary: cleaning, cooking, laundry, setting a table, moving bulky objects. The phrase "open-world" in the paper's title points at exactly this spread of places and tasks. The authors are listed as the Lambda-humanoid team, but the affiliation reads Token Hub and Amap at Alibaba Group, and collection was run by Alibaba's data technology team together with Amap's embodied AI data team. An earlier paper introducing the same dataset and model went up on 30 September 2026, and the author names appear in full there.

This team is not the only one looking in that direction. The related-work section lists several attempts from the same period. EgoVLA trained a vision-language-action model on first-person video, HALOMI carried human demonstrations as far as active perception, and EMMA trained human whole-body loco-manipulation demonstrations alongside data from a fixed-base robot. The authors note that EMMA also reported scores rising as human data grew, and place their own difference in using body and hand motion together as a supervision signal for loco-manipulation on a bipedal humanoid. Moving the bottleneck toward people is under way in several places at once.

Human video cannot be fed to a robot as it stands. The lower part of the frame shows a person's arms and hands, and the robot does not have those arms. Height and arm length differ, so the same posture puts the fingertips somewhere else. Joints do not follow a command exactly either. That is why the researchers split the move from human to humanoid into four stages.

2.1Matching What the Camera Sees

The first stage repairs the picture itself. SAM2 segments the human arms out of the video, DiffuEraser paints over the place where they were, and a rendering of the G1 arms at the same joint positions is composited into the gap. The work makes the scene the trained model will meet on the real robot and the scene it saw during training look alike.

2.2Matching the Movement

Next comes the body. The human body model (SMPL) is scaled down to the G1's height and the pose is corrected, then the motion is rolled out once on the robot body through the SONIC whole-body controller. The error left over by the difference in skeletons is shaved again with constrained offline whole-body inverse kinematics. Finally the mismatch produced during actual execution is fed back, so the motion signal handed to the robot is itself adjusted step by step.

The effect of all four stages was measured on a validation set of 50 tasks running about an hour. Task-average fingertip position error fell from 17.03cm to 3.83cm, and orientation error from 36.06 degrees to 10.85 degrees. 17cm is the distance that brushes past the handle of a cup, and 3.8cm is about two finger joints.

The model trained on this data is λ₀. Training splits into three stages. Stage one is pre-training on seven public egocentric datasets, EgoDex and HOI4D among them, predicting the state of both hands. Stage two is mid-training, where HumanVerse is converted into a form the robot can read and whole-body coordination is learned. Stage three is post-training on real robot demonstrations. The vision-language model at the core is Qwen3.5-2B, and the output is passed on the G1 to the SONIC whole-body controller and BrainCo Revo2 hands.

In simulation, λ₀ recorded 90.0% on SIMPLE, a benchmark the team defined itself. That figure comes from six tasks split across three difficulty levels from 0 to 2, ten trials for every combination, 180 trials in all. Under the same conditions StarVLA scored 86.1%, Ψ₀ 85.6%, π₀.₅ 51.1% and GR00T N1.6 48.9%. SIMPLE is worth keeping apart from the similarly named SIMPLER environment, since this one belongs to the team.

On the physical Unitree G1, four tasks were run ten times each, 40 trials in all: throwing away a bottle, tidying toys, moving a box and placing a chair. The progress rate in the table adds up credit for checkpoints fixed in advance for each task, converts the result to a 100-point scale and averages the tasks evenly.

▲ Table 1 of the paper. Each cell gives the number of successes out of ten trials and the progress rate. By successes π₀.₅ stands highest among the baselines, by progress rate StarVLA does.

λ₀ succeeded on 25 of 40 trials, 62.5%, with a progress rate of 80.4%. The baseline with the most successes is π₀.₅ at 16 trials (40.0%), and the baseline with the highest progress rate is StarVLA at 54.6%. By either measure the gap is not small.

3.1What Is Left When the Human Video Is Taken Out

How much of that score belongs to HumanVerse was measured on its own. The architecture and the robot post-training were held fixed while training stages were dropped one at a time. Removing stage-two mid-training takes the real-world success rate from 62.5% to 22.5% and the progress rate from 80.4% to 59.9%. That is a far steeper fall than removing stage-one pre-training, which leaves 50.0% and 75.9%. Dropping both leaves 15.0% and 54.0%.

A curve measured by raising the volume in steps points the same way. Changing the share of HumanVerse fed into stage two across six settings from 0% to 100% moved the real-world progress rate to 59.9%, 64.2%, 67.3%, 72.1%, 76.4% and 80.4%, climbing at every step without once falling back. End to end that is a spread of 20.5 points.

3.2"Without Robot Demonstrations" Points at a Different Experiment

This is where summaries of the paper blur together most often. The abstract says the work transfers to a Unitree G1 "without target-task robot demonstrations". That sentence refers to the pipeline validation of section 5.1. Four whole-body tasks, closing a laptop, opening a fridge, pulling a curtain and closing a drawer, and three long-horizon tasks, object relocation, navigation and foot interaction, were carried straight over to the robot from aligned human demonstrations alone, at average progress rates of 51.7% and 76.2%.

The 62.5% in the table above belongs to a different experiment. The λ₀ of section 5.2 went through stage-three robot post-training, and the four baselines it is set against were evaluated under the same conditions. Joining the two experiments into "62.5% without a single robot demonstration" claims one step more than the paper does. Put precisely: human demonstrations alone carry a robot through more than half of a task, and a model with robot data layered on top does far better than a model trained on that robot data alone.

The researchers also checked whether human video fills the gaps in robot data. Tested on objects that appear only in the human video and never in the robot demonstrations, λ₀'s progress rate rose to 55.8% from the 48.4% of the same model trained without human data. It rose on all four tasks, though it does not reach the scores for objects the robot has already seen. The researchers add that the comparison takes the effect of the human data as a whole and cannot isolate the contribution of footage of any particular object.

That is the record of what the paper reports. Anyone who sets out to check the result runs into a wall soon enough. The OpenLambda code repository is published under Apache 2.0, and the model checkpoint from stage-two mid-training can be downloaded from Hugging Face. What cannot be downloaded is the data that produced that checkpoint. The repository states that "HumanVerse-500 training data is not included in this code release", and what can be fetched right now is four motion-alignment examples and sample trajectories for the four tasks. The datasets used for pre-training and mid-training fall outside the release as well. The project page links to a sample dataset on Hugging Face, but as of 16 October 2026 that address does not open from outside and returns an authorization error.

The code that is public is not all of it yet either. The repository writes its remaining work down plainly. There is no recorder yet for capturing episodes with the G1, the SONIC controller, BrainCo hands and the head camera, and no converter that turns human records and robot demonstrations into λ₀'s standard episode format. Where the visual alignment code from section 2 would sit, the entry reads "coming soon", and the simulation evaluation and real-robot deployment guides are in preparation. Stage-one and stage-two training code and configs are up, though the repository itself says it does not guarantee end-to-end reproduction from them.

What the "open world" in the title opens is the location of the filming, not the data. The model can be downloaded and run, while what it learned from stays as prose in the paper. From outside there is now no way to count whether 500 hours is really 500 hours, whether the spread of places matches the description, or whether the footage leans toward particular body types or ways of living.

The handling of the data itself is written down carefully enough. According to the appendix, every episode carries a fixed specification recording its source, capture metadata, annotation coverage, conversion version and split assignment. Timestamps, stream overlap, coordinate conventions and valid ranges are checked structurally, and hands are reprojected onto the video for inspection by eye. Splits are drawn by capture session, so scenes from one session do not land in training and validation together. Task-specific human demonstrations gathered for post-training are kept apart from the general mid-training corpus.

What is missing sits on the other side. Nowhere in the body of the paper, the appendix, the project page or the code repository is there a sentence about participant consent, ethics review, or the handling of faces and voices. Sixteen data collectors are named in the acknowledgements, so this is plainly a record a team made with people in front of them rather than video scraped off the internet. An internal procedure most likely existed, but from the published documents alone there is no telling what it was.

The bottleneck in physical AI has sat on the robot side for a long time. Robots have to exist for data to pile up, and data has to pile up for robots to multiply. Where do you break that circle? This blog has sorted robot learning data into five layers before, and has covered the gap where tactile data stays empty while capital pours in. This research is an attempt to break that circle on the human side. Once the bottleneck changed places, the questions changed with it.

There are three such questions. First, there is still the job of turning filmed footage into training data a robot can use. The four stages that took the fingertip error from 17cm to 3.8cm are that job. As collection grows cheaper, processing and the work of verifying that processing grow more expensive. Second, the paper shows in concrete terms what has to be recorded alongside human video for the video to be of use. Video alone falls short: whole-body pose, hand motion, capture clock, session identifier and converter version have to travel as one bundle. Third, the bundle holds no entry about the person.

For a company that third item surfaces last and costs the most. Plans to film employees at work in first person in a plant or a store and feed it to robot training are already under review in several places. Unless the episode specification also records who was filming, what consent they gave and how long the footage may be kept, the question of whether the data can be used at all comes back once the model is finished. Even a team that keeps a meticulous processing history will lose the record about people unless it is raised as an entry of its own.

It is hard to fault the choice not to publish the data. Releasing 500 hours of video showing people's homes and faces would create the larger problem. Still, if the reason for keeping it closed is privacy, writing that reason down is the better move. Right now there is the fact of non-release and no account of why, which leaves readers to fill the gap with guesses.

Thank you for reading this far. The full paper is on arXiv, and the code is in the OpenLambda repository. If your team is piling up data that records people, how many entries in that record are about the people themselves? If you have ever counted, we would be glad to see the list.

• 1.Lambda-humanoid Team. (2026). "𝜆₀: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation." arXiv:2610.18950.
• 2.Lambda-humanoid Team. (2026). "OpenLambda." GitHub repository.
• 3.Lambda-humanoid Team. (2026). "λ₀ project page."
• 4.Lambda-humanoid Team. (2026). "lambda0-midtrain-100k." Hugging Face.