Humanoid robot videos usually show one trick at a time: a backflip in one clip, a dance routine in the next, each tuned by engineers until it looks right. The research behind Tech Xplore’s 10 September 2026 story, “Humanoid robot learns to sprint and perform spin kicks using AI trained on human motion data”, matters for the opposite reason. One training recipe, with one set of settings, taught a Unitree G1 humanoid robot to sprint, throw spin kicks, crawl, dance and land aerial cartwheels.
The framework is called BeyondMimic. Researchers at the University of California, Berkeley and Stanford University built it, and Science Robotics published it on 26 August 2026 (DOI 10.1126/scirobotics.adx8924). The coverage leans on three figures: about 2.5 hours of human motion, a 77-person user study, and a 70.8% share of comparisons in which viewers judged the humanoid robot’s movement more humanlike than motion from Unitree’s own controller.
We checked those figures against the open preprint, the project page, the code and the robot’s store listing. They hold up, but each needs context. The 70.8% blends a near coin flip on walking with a landslide on running. The hardware tests covered a tenth of the motion used in training. The obstacle demonstrations relied on a motion capture system. And the code has been public for 13 months, long enough to be running on other robots. For more research coverage, see our AI Models, Tools and Releases hub.
Table of contents
- What the Humanoid Robot Actually Learned to Do
- How BeyondMimic Turns Human Motion Data Into Humanoid Robot Skills
- The Diffusion Model That Chains Humanoid Robot Moves Together
- Is the Humanoid Robot Really More Humanlike? Splitting the 70.8%
- What Ran on the Real Humanoid Robot and What Stayed in Simulation
- What the Humanoid Robot Demos Leave Out
- The Humanoid Robot Hardware Behind the Demo: Unitree G1
- An Open-Source Recipe Other Humanoid Robot Makers Already Use
- Sprint Records at the World Humanoid Robot Games Are a Different Race
- What This Humanoid Robot Research Means for Businesses
- Humanoid Robot FAQ: BeyondMimic in Brief
- References and Further Reading
What the Humanoid Robot Actually Learned to Do
The skill list reads like a gymnastics syllabus. Tracking policies deployed on the real humanoid robot covered balance moves such as single-leg standing and getting up from different poses; dynamic skills such as single-leg jumps, turn kicks, forward jumps with 180° and 360° spins, and cartwheels; and stylised movement, including an elderly-style walk, dance sequences and sport moves. One clip reproduces Cristiano Ronaldo’s celebration jump-turn.
Sprinting, spin kicks and cartwheels from one recipe
No single move on that list is a first for a humanoid robot. What is new is how the moves were obtained. The journal abstract says earlier methods “either produced unnatural motions or relied on motion-specific tuning” to look natural, and that they were usually built for one motion or one goal. BeyondMimic’s claim is that one compact tracking formulation handles aerial cartwheels, spin kicks, flip kicks and sprinting “with a single setup”, using the same reward design and training settings for every clip.
That matters more than any single stunt. If a new skill needs a new recording rather than a new reward function, the work of teaching a humanoid robot another movement shifts from controller engineering toward collecting and cleaning motion recordings.
Who built BeyondMimic and where it was published
The Science Robotics paper lists seven authors. Qiayuan Liao, Xiaoyu Huang, Yuman Gao and Koushil Sreenath are at UC Berkeley; Takara E. Truong, Guy Tevet and C. Karen Liu are at Stanford. The project page marks Liao, Truong and Huang as equal contributors, with their order “decided by coin toss”, and lists Sreenath and Liu as equal advisers. The code lives under the HybridRobotics organisation on GitHub, which shares its name with Sreenath’s Hybrid Robotics group at Berkeley.
The preprint credits funding from the US National Science Foundation, the TIAMAT programme of the US Defense Advanced Research Projects Agency, the Robotics and AI Institute, the BAIR Humanoid Intelligence Center, Stanford’s Institute for Human-Centered AI and the Wu-Tsai Human Performance Alliance. It also thanks Amazon for the use of its facility.
| Item | Detail |
|---|---|
| Framework | BeyondMimic: motion tracking plus guided latent diffusion |
| Institutions | University of California, Berkeley and Stanford University |
| Journal | Science Robotics, volume 11, issue 117, article eadx8924 |
| Journal publication | 26 August 2026 |
| First preprint | arXiv 2508.08241, 11 August 2025 (version 4 on 13 November 2025) |
| Robot | Unitree G1 |
| Human motion used | About 2.5 hours |
| Tested on hardware | 30 clips totalling 15 minutes (arXiv version 4) |
| User study | 77 participants, 20 paired clips each |
| Code | HybridRobotics/whole_body_tracking on GitHub, MIT licence |
A 13-month-old preprint behind a September headline
BeyondMimic is not brand new. Version one of the preprint went up on arXiv on 11 August 2025, the same day the tracking code repository was created on GitHub, and version four followed on 13 November 2025. The peer-reviewed journal version arrived 380 days after the first preprint. Interesting Engineering covered it on publication day, a KameraOne video syndicated by Yahoo News UK followed on 2 September, and Tech Xplore’s story by Ingrid Fadelli ran 15 days after publication.
One detail shows how the humanoid robot project grew in that time. The project page still lists six authors, while the latest arXiv version and the journal list seven, adding Yuman Gao.
How BeyondMimic Turns Human Motion Data Into Humanoid Robot Skills
BeyondMimic has two stages, and it helps to keep them apart, because news coverage often blends them. The first stage teaches the humanoid robot to copy individual motions accurately. The second, covered in the next section, learns from those copies and composes them into new behaviour.
Stage one: a tracking policy for each motion
The first stage uses reinforcement learning, the trial-and-error method in which a controller is rewarded for good outcomes and penalised for bad ones. Here the reward is for tracking: matching the positions, orientations and speeds of the robot’s body parts to the reference motion at each moment. According to Tech Xplore’s summary, penalties discourage jerky movements, unsafe joint positions and unwanted contact between the humanoid robot’s own body parts.
The algorithm is PPO, one of the most widely used methods in robot learning; our guide to Proximal Policy Optimization explains how it works. The policies themselves are small: multilayer perceptrons with hidden layers of 512, 256 and 128 units for both the actor and the critic, running at 50 Hz on the robot.
One reward design and one set of settings for every clip
The supplementary table of training settings is short, and that is the point. Every tracking policy uses a learning rate of 0.001, a discount factor of 0.99, a PPO clip parameter of 0.2 and a cap of 30,000 training iterations. The authors say the same model parameters, reward function and settings carry across a diverse set of skills and different physics engines. That is what lets the humanoid robot pick up a new motion without a fresh round of reward engineering.
Adaptive sampling spends practice on the hardest seconds
Long reference clips mix easy walking with a few brutal seconds, such as a cartwheel. Sampling every moment equally wastes practice on the easy parts. BeyondMimic’s adaptive sampling instead prioritises segments where the policy fails most often, then drifts back toward uniform sampling once those segments are mastered.
The ablation is stark. Without adaptive sampling, three of four test motions still failed in challenging segments after 30,000 iterations, including the cartwheels in one clip. Even on an easy motion, adaptive sampling halved the iterations needed, from 4,000 to 2,000.
Where the 2.5 hours of human motion came from
The paper says the team trained on approximately 2.5 hours of diverse human motion and validated every motion in high-fidelity simulation. That motion came from several places: a version of Ubisoft La Forge’s LAFAN1 motion-capture dataset retargeted to the G1 by Unitree, motions released with three earlier humanoid robot papers (ASAP, HuB and KungfuBot), and two animation packs from Reallusion’s online stores, “MD Panther Lady” and “Martial Arts – Taekwondo”.
Retargeting is the unglamorous step. Human recordings have to be mapped onto a robot with different proportions, joint limits and mass, so the training data here is motion already adapted to the G1’s body rather than raw human recordings. Other groups attack the same shortage from the simulation side, as our report on AI agents that build virtual playgrounds for robots describes.
| Source | What it is | Licence or terms | Used for |
|---|---|---|---|
| LAFAN1 (Ubisoft La Forge), retargeted by Unitree | Studio motion capture: 5 subjects, 77 sequences, about 4.6 hours at 30 fps | CC BY-NC-ND 4.0 | Walking, running, dancing, falling and getting up |
| ASAP (2025 paper) | Motions released with an earlier agile-control paper | Set by the original release | Cristiano Ronaldo celebration |
| HuB (2025 paper) | Extreme balance motions | Set by the original release | Single-leg and swallow balances |
| KungfuBot (2025 paper) | Highly dynamic martial-arts motions | Set by the original release | Side kick |
| Reallusion “MD Panther Lady” pack | iClone animation pack from Reallusion’s content store | Store licence terms | Stylised movement |
| ActorCore “Martial Arts – Taekwondo” pack | Motion pack from Reallusion’s ActorCore store | Store licence terms | Martial-arts movement |
LAFAN1 alone holds 77 sequences from five subjects, about 4.6 hours at 30 frames per second. The project page shows policies trained on 14 of those sequences, each roughly three minutes long, so around 42 minutes of LAFAN1 appears in that showcase.
The Diffusion Model That Chains Humanoid Robot Moves Together
A library of tracking policies can replay motions, but it cannot decide what to do next. BeyondMimic’s second stage distils the tracked skills into one model that can generate new movement toward goals it never saw during training.
Compressing movement into a compact code
First, a variational autoencoder compresses the robot’s actions into a compact latent representation, and a lightweight decoder turns that representation back into motor commands. The team doubled the skill set cheaply by also training on mirror-image copies of every motion, with left and right swapped.
Then a diffusion model, the same family of technique behind many image generators, learns sequences that combine the humanoid robot’s predicted physical states with those compressed actions. Training is self-supervised: noise is added to real trajectories, and the model learns to recover the clean version.
Classifier guidance steers the humanoid robot at test time
Guidance is the feature that takes the framework beyond mimicry. Because a diffusion model refines a noisy guess step by step, a cost function can nudge each step toward a goal: follow this velocity, reach this waypoint, keep clear of that obstacle. The paper calls this classifier guidance, a technique first introduced for image generation, and uses it as online optimisation. A new task needs a new cost function, not a new training run.
Joystick, waypoints, obstacles and keyframes
The demonstrations cover four kinds of goal. With joystick commands, the humanoid robot walked in any direction and tracked the commanded heading. With waypoints, it produced smooth paths to targets from different starting positions. With a signed distance field describing obstacles, it detoured around them while still heading for its goal. With motion inpainting, where a few future keyframes are specified, it filled in the movement between them: walking into a cartwheel and back, or three consecutive cartwheels stitched together.
The latent step is what made the agile moves reliable. In a MuJoCo sim-to-sim test of aerial cartwheels, a baseline without the latent encoding completed the move without falling only 5% of the time. The latent diffusion model managed 95%, and the authors report that the performance carried over to the physical humanoid robot.
| Aspect | Stage one: motion tracking | Stage two: guided latent diffusion |
|---|---|---|
| What it learns | To follow a reference motion closely, including clips over three minutes long | Sequences of robot states and compressed actions across many motions |
| How it is trained | PPO with one shared reward design and set of settings | Autoencoder compression, then self-supervised denoising |
| Where it runs | 50 Hz policy in C++ on the robot | About 20 ms per inference step on an onboard RTX 4060 Mobile |
| New goals at test time | No, it follows its reference | Yes, through cost functions and classifier guidance |
| Known weak spots | Cannot choose or combine skills on its own | 0.64-second horizon; stumbles at the start and end of motions |
Is the Humanoid Robot Really More Humanlike? Splitting the 70.8%
Every article about BeyondMimic quotes the same number: in 70.8% of comparisons, people judged its movement more humanlike and natural than Unitree’s standard controller. The figure is accurate. It is also an average of two very different results.
How the user study worked
The study had 77 participants. Each viewed 20 pairs of five-second clips of walking and running, one clip from BeyondMimic and one from Unitree’s native controller, which the authors describe as the state of the art for this humanoid robot. Participants answered one question: “Which looks more human-like and natural?” The team used a two-tailed binomial test, with a Bonferroni correction for the separate walking and running comparisons.
Running carried the headline number
Split by gait, BeyondMimic’s running was preferred in 84.7% of choices and Unitree’s in 15.3%. For walking, the split was 57.0% to 43.0%. Both preferences were statistically significant, but they are not the same kind of win. Running was a landslide; walking was much closer to a coin flip.
The arithmetic behind 70.8%
Three calculations put the headline in context. First, 77 participants each judging 20 pairs gives 1,540 judgements. Second, the overall 70.8% sits almost exactly halfway between the two gaits, since (57.0 + 84.7) ÷ 2 = 70.85, which is what an equal mix of walking and running pairs would produce. Third, against the 50% a random choice would give, walking was 7.0 percentage points above chance and running 34.7 points above it, roughly five times the margin.
If you are assessing a humanoid robot for work that is mostly walking, such as moving along warehouse aisles or guiding visitors, the 57.0% is the more relevant number, not the 70.8%.
A missing toe joint shows up in walking
The paper offers a mechanical explanation. The team compared ground reaction forces from the robot with human data recorded on force-sensing treadmills, and the shapes matched well, including the double peak of a heel strike followed by push-off. In walking, though, the humanoid robot showed sharper force peaks, which the authors attribute to the G1’s lack of a toe joint limiting roll-over and push-off. In running, each step makes a single, shorter contact, so the missing toe matters less.
What Ran on the Real Humanoid Robot and What Stayed in Simulation
Simulation results are cheap; hardware results are the ones that count. BeyondMimic has both, in very different quantities.
30 clips and 15 minutes on hardware
All of the roughly 2.5 hours of motion was validated in simulation. To test transfer to the real world, the team deployed 30 representative clips, totalling 15 minutes, on the physical humanoid robot, according to the latest arXiv version. That is 15 of about 150 minutes, or 10% of the motion used in training.
Ten per cent is not a flaw in itself. Hardware time is expensive, falls damage robots, and testing a representative sample is normal practice. It does mean that the paper’s description of a framework that learns “hundreds of diverse motions” is a simulation claim, while the real-world evidence rests on a curated subset.
Soft soil, decaying leaves and uneven ground
The outdoor tests were deliberately messy. The authors evaluated agility on soft soil, decaying leaves and uneven ground, conditions that introduce deformable and unstable contacts not present in training. The humanoid robot still completed acrobatic sequences and martial-arts-inspired motions, and the paper also shows contact-rich skills such as two consecutive cartwheels, crawling and jumping up from the ground.
Cartwheel physics compared with skilled humans
During the airborne phase of an aerial cartwheel, the robot reached a peak acceleration of 31 m/s² and a pelvic angular velocity of up to 20 rad/s, averaging 7.01 rad/s. The paper sets that against an average of 7.75 rad/s reported for skilled human aerial movements, so the humanoid robot’s mean rotation rate was about 90% of the human figure. Its landings needed minimal recovery.
Where the coverage and the preprint disagree
Reading the reports side by side turns up discrepancies. We could not check them against the final journal text, because Science’s website blocks automated access, so the table compares the latest arXiv version with the two most detailed news reports. Interesting Engineering, which reported on publication day, may reflect figures revised during peer review; Tech Xplore’s clip count matches the preprint.
| Figure | arXiv version 4 (13 Nov 2025) | Interesting Engineering (26 Aug 2026) | Tech Xplore (10 Sep 2026) |
|---|---|---|---|
| Human motion used | About 2.5 hours | About 2.5 hours | About 2.5 hours |
| Clips deployed on hardware | 30, totalling 15 minutes | 21 | 30 |
| Peak pelvic angular velocity in a cartwheel | Up to 20 rad/s (mean 7.01) | Up to 15.7 rad/s | Not reported |
| Cartwheels chained in a guided sequence | Three consecutive | Four, with walking and running | Not reported |
| User study | 77 participants, 70.8% overall | 77 participants, 70.8% | 77 participants, 70.8% |
| Walking and running split | 57.0% and 84.7% | Not reported | Not reported |
The practical lesson is to cite the version you actually read. A preprint and its journal article are separate documents, and a figure such as a clip count can differ between them.
What the Humanoid Robot Demos Leave Out
The methods and limitations sections of the paper are more candid than the headlines. Five details change how the demonstrations should be read.
Motion capture placed the waypoints and the obstacles
For basic motion tracking, the robot used no external motion capture. It estimated its own state at 500 Hz with a generalised momentum observer and a Kalman filter. For the hardest contact-rich moves, such as getting up from the ground, the team either added LiDAR-inertial odometry for position correction or dropped the observations that depend on state estimation.
The navigation demos were different. For waypoint navigation and obstacle avoidance, the supplementary material says motion capture data supplied both the environmental context for the cost functions and improved localisation, and the project page caption says the same. In plain terms, the humanoid robot did not see the obstacle it avoided; a motion capture system told the controller where it was. That is a sound way to test a controller, but it is not a deployable perception system.
A 0.64-second view of the future
The diffusion model predicts trajectories 0.64 seconds ahead. The authors say that is enough for reactive control and local obstacle avoidance but not for long-horizon planning, where the robot must reason about distant goals or anticipate obstacles early.
Stumbles at the start and end of guided motions
Guidance has a side effect. Feeding the model recent history stabilises its predictions but can trap it in repetitive patterns, so the team raised the guidance weights, which can in turn destabilise denoising when the humanoid robot switches modes. The paper’s own summary is blunt: “under guided diffusion, the robot is stable once a gait orbit is established, but tends to stumble at the start and end of motions.”
Coarse goals work better than fine ones
Guidance works well for coarse objectives, such as a heading or a waypoint, and less well for fine-grained ones. It still needs light tuning of guidance weights, and errors in the robot’s own state estimate flow straight into the trajectories it generates.
Whole-body movement, not object handling
Every demonstration in the paper is whole-body movement: walking, running, acrobatics, getting up and navigating. None involves grasping, carrying or using an object, which is where much of the commercial interest in humanoid robots lies.
| Demonstration | What it shows | What the methods section adds |
|---|---|---|
| Outdoor motion tracking | Cartwheels and martial-arts moves on soil and leaves | No external motion capture; onboard state estimation at 500 Hz |
| Getting up from the ground | Contact-rich recovery | LiDAR-inertial odometry for position correction, or no state-estimate inputs |
| Joystick control | Walking in any commanded direction | Onboard proprioceptive state estimation only |
| Waypoint navigation | Smooth paths to set targets | Motion capture supplies targets and localisation |
| Obstacle avoidance | Detours around obstacles | Motion capture supplies obstacle positions for the cost function |
| Keyframe transitions | Walking into cartwheels and back | Stable once a gait is established; tends to stumble at the start and end of motions |
The Humanoid Robot Hardware Behind the Demo: Unitree G1
BeyondMimic ran on Unitree’s G1, a compact humanoid that has become a common research platform; both MJLab and Unitree’s own RL Lab ship G1 training tasks. Two details from the paper and the store listing matter for anyone costing a replication.
A $13,500 list price that is not the research price
Unitree’s online store lists the G1 at $13,500, plus shipping of $300 to $1,200 and any customs duties, and showed it as backordered on 10 September 2026. The listing carries a warning: “This product does not support secondary development. For customization needs, please select the EDU edition.” The EDU edition is sold through Unitree’s sales team, with no public price.
In other words, the headline price buys a humanoid robot you are not meant to load your own controller onto. A lab or business reproducing BeyondMimic needs the development edition. The paper’s supplementary material also computes actuator parameters for four G1 robots, which suggests the team worked across several units. The store lists the G1 with 23 to 43 joints.
The computer riding on the robot
The tracking policies run at 50 Hz in C++ code written for real-time execution. The diffusion policy runs onboard, on a portable mini PC with an NVIDIA RTX 4060 Mobile GPU, accelerated with TensorRT. Each diffusion inference step takes about 20 ms using 20 denoising steps and runs asynchronously in its own thread, while the lightweight latent decoder runs synchronously on the CPU.
That is laptop-class gaming hardware rather than a data-centre GPU, which is encouraging for anyone pricing a humanoid robot deployment. For a comparison with embedded modules designed for robots, see our look at Jetson Orin Nano 2 for physical AI.
| Layer | Detail | Source |
|---|---|---|
| Robot | Unitree G1; actuator values computed for four units | Paper supplement |
| Listed price | $13,500 plus $300 to $1,200 shipping | Unitree store |
| Development access | Standard G1 does not support secondary development; EDU edition through sales | Unitree store |
| Joints | 23 to 43 | Unitree store |
| Tracking policy | 50 Hz, C++ deployment code | Paper supplement |
| State estimation | 500 Hz momentum observer with a Kalman filter | Paper supplement |
| Diffusion inference | Mini PC with RTX 4060 Mobile, TensorRT, about 20 ms per step with 20 denoising steps | Paper supplement |
| Latent decoder | Runs synchronously on the CPU | Paper supplement |
An Open-Source Recipe Other Humanoid Robot Makers Already Use
The strongest evidence that BeyondMimic works is not in the journal. It is on GitHub.
2,386 stars in 13 months
The HybridRobotics whole_body_tracking repository, which holds the motion tracking code, was created on 11 August 2025 under the MIT licence. On 10 September 2026 it had 2,386 stars and 326 forks. A GitHub search for “beyondmimic” returned 73 repositories, including deployment tools, reproduction guides and ports to other humanoid robot hardware.
Built into MJLab and Unitree RL Lab
The paper says BeyondMimic’s open-source release now serves as a default method in public repositories such as MJLab and Unitree RL Lab. Both check out. MJLab, an Apache-2.0 framework with 3,032 stars, includes a motion tracking task for the G1 and describes it as a way to “Train a humanoid to mimic reference motions.” Unitree’s own RL Lab ships “mimic” deployment policies for its 29-joint G1, including a dance routine and a Gangnam Style routine.
A second robot maker’s port
Adoption is not limited to Unitree hardware. HighTorque Robotics published Mini-Pi-Plus_BeyondMimic, an MIT-licensed port named for its Mini Pi Plus humanoid robot, on 17 October 2025. Community projects such as Beyondmimic_Deploy_G1 package the pipeline for G1 owners. MJLab and Unitree RL Lab are general frameworks that predate BeyondMimic; the other repositories below exist because of it.
| Repository | Maintainer | Created | Licence | Stars (10 Sep 2026) |
|---|---|---|---|---|
| whole_body_tracking | HybridRobotics | 11 Aug 2025 | MIT | 2,386 |
| mjlab | mujocolab | 10 Jun 2025 | Apache-2.0 | 3,032 |
| unitree_rl_lab | Unitree Robotics | 5 Jun 2025 | Apache-2.0 | 1,336 |
| Beyondmimic_Deploy_G1 | Ming827 (community) | 2 Sep 2025 | Not stated | 193 |
| Mini-Pi-Plus_BeyondMimic | HighTorque Robotics | 17 Oct 2025 | MIT | 182 |
| BeyondMimic_G1 | Xingzb (community) | 19 Dec 2025 | MIT | 48 |
The code is MIT; the motion data is not
The permissive code licence does not cover the motion the robot learned from. LAFAN1 is released under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0, which rules out commercial use, and the animation packs carry their stores’ own licence terms. A business that wants a humanoid robot to perform in a commercial setting should budget for motion it has the right to use, whether licensed or captured in-house, and take legal advice before training on retargeted copies of research datasets.
Sprint Records at the World Humanoid Robot Games Are a Different Race
BeyondMimic’s sprint reached the news two weeks after a very different one. At the second World Humanoid Robot Games in Beijing, a humanoid robot ran 100 metres faster than any human has.
Tiangong Ultra went from 9.39 to 8.64 seconds
The Games ran from 22 to 26 August 2026 at Beijing’s National Speed Skating Oval, with 666 teams and 2,056 robots from 16 countries across 51 events, according to the Philippine Canadian Inquirer. Tiangong Ultra, built by the Beijing Humanoid Robot Innovation Center, ran 9.39 seconds early in the Games, then 8.86 seconds, then 8.64 seconds in the final. Usain Bolt’s world record, set in 2009, is 9.58 seconds.
Tech Times reported that Honor’s robot Lightning ran 9.47 seconds in the same opening-day heat as Tiangong Ultra’s 9.39, that both were unable to brake in time and ran into the padded buffer wall at the end of the track, and that the winning time at the inaugural Games a year earlier was 21.50 seconds. Going from 21.50 to 8.64 seconds is a 59.8% cut in one year.
Speed-first training can find non-human answers
The sharper contrast is in technique. Tech Times, citing the Global Times, reported that an X-Humanoid robot abandoned the humanlike arm swing its engineers designed for its 400-metre run. It ran with its arms held near its face, weight pitched forward and hips driving the stride, a technique X-Humanoid motion control engineer Han Gang said emerged from reinforcement learning in simulation.
That is the opposite design goal to BeyondMimic. A robot optimised for race time will settle on whatever gait is fastest, human or not. BeyondMimic rewards the humanoid robot for matching recorded human motion, so its sprint is judged on whether it looks like a person running. The paper publishes no top speed or 100-metre time, so the two cannot be compared on pace.
| Aspect | BeyondMimic sprint | World Humanoid Robot Games 100 m |
|---|---|---|
| Goal | Look like a person sprinting | Cover 100 metres as fast as possible |
| What shapes the gait | A reward for matching recorded human motion | Speed; one reported X-Humanoid technique emerged in simulation training |
| Robots | Unitree G1 | Tiangong Ultra, Honor’s Lightning and others |
| Headline result | Running preferred in 84.7% of choices | 8.64 seconds in the final |
| How it is measured | A 77-person preference study | Race timing |
| Stopping | Guided transitions back into walking | Two opening-day sprinters hit the buffer wall |
What This Humanoid Robot Research Means for Businesses
Few businesses will buy a robot to do cartwheels. The recipe behind the cartwheels, though, addresses a real bottleneck in humanoid robot deployment: every new movement has traditionally meant new controller work.
New skills become a data problem
If one training set-up covers sprinting, kicking and crawling, then adding a task-specific movement, such as a particular way of stepping around shelving, becomes a matter of recording and retargeting motion rather than designing a new controller. That moves cost away from specialist control engineering and toward motion capture, data cleaning and simulation time.
Put motion data rights in the budget
The licence gap described above is the easiest part to miss. Code licences are visible on GitHub; dataset and animation licences are not always read. A commercial humanoid robot programme should treat motion data like any other licensed content, with the source and terms recorded for every clip.
Ask for split results, not blended ones
The user study shows why averages mislead. A vendor quoting a single humanlike or success score should be asked for the breakdown by task, because a 57.0% walking result and an 84.7% running result can produce the same 70.8% headline.
Test in simulation, then ask what the demo relied on
BeyondMimic validated every motion in simulation before touching hardware, a workflow that tools such as Antioch’s browser-based robotics simulation aim to make routine. When you watch any humanoid robot demo, ask what sat off camera: motion capture, a safety tether, or a curated set of takes. For how a large manufacturer frames the opportunity, see our analysis of Hyundai Robotics and its physical AI plans.
| Question to ask | Why it matters | What the BeyondMimic paper shows |
|---|---|---|
| How much was tested on real hardware? | Simulation success does not guarantee real-world success | 30 clips and 15 minutes, out of about 2.5 hours |
| Are results split by task? | Averages hide weak spots | Walking 57.0%, running 84.7%, overall 70.8% |
| What off-board equipment did the demo use? | Motion capture does not ship with the robot | Waypoints and obstacles came from motion capture |
| Who owns the motion recordings? | Licences can block commercial use | LAFAN1 is CC BY-NC-ND 4.0 |
| Which hardware edition is needed? | Retail units may be closed to custom code | The standard G1 does not support secondary development |
| How far ahead does the controller plan? | Short horizons limit navigation | 0.64 seconds |
Humanoid Robot FAQ: BeyondMimic in Brief
What is BeyondMimic?
BeyondMimic is a framework from UC Berkeley and Stanford for teaching a humanoid robot agile, humanlike skills from human motion. It combines motion tracking policies trained with PPO and a latent diffusion model that can be steered toward new goals at test time.
Which humanoid robot did the researchers use?
The Unitree G1. Unitree’s store lists the standard G1 at $13,500 but says that version does not support secondary development, which requires the EDU edition.
What human motion data was it trained on?
About 2.5 hours of motion, drawn from a retargeted version of Ubisoft La Forge’s LAFAN1 motion-capture dataset, motions from the ASAP, HuB and KungfuBot research papers, and two Reallusion animation packs.
Can businesses use the BeyondMimic code?
The motion tracking code is on GitHub under the MIT licence. The motion data is a separate question: LAFAN1’s licence is non-commercial, so check the terms of any dataset before commercial use.
Is it faster than the robots that beat Usain Bolt’s record?
The comparison does not work. BeyondMimic’s sprint is judged on how humanlike it looks, and the paper reports no race time. Tiangong Ultra ran 8.64 seconds in the 100-metre final at the 2026 World Humanoid Robot Games.
When was the research published?
The first arXiv preprint appeared on 11 August 2025. Science Robotics published the peer-reviewed paper on 26 August 2026, and Tech Xplore reported on it on 10 September 2026.
References and Further Reading
BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion (arXiv)
HybridRobotics whole_body_tracking repository
New robot system lets humanoids learn cartwheels, kicks and more (Interesting Engineering)
Humanoid robot does cartwheels and flip kicks like a human (KameraOne via Yahoo News UK)
Ubisoft La Forge Animation Dataset (LAFAN1)
HighTorque Mini-Pi-Plus_BeyondMimic repository
HuB: Learning Extreme Humanoid Balance (arXiv)
KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills (arXiv)
Proximal Policy Optimization Algorithms (arXiv)
AI-Trained Robot Invented Its Own Sprint at 2026 World Humanoid Robot Games (Tech Times)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.