AnswersPublished19 sources
Being-M0.7 uses human motion data to teach humanoid robots to move and manipulate
Being M0.7 is a latent world action model designed to coordinate humanoid walking and object manipulation, rather than training them as separate skills.[1][6] It learns from more than 10,000 hours of mixed human video and motion data, then uses robot trajectories to adapt its predictions into whole body commands.[1]...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Being M0.7 is a latent world action model designed to coordinate humanoid walking and object manipulation, rather than training them as separate skills.[1][6]
- It learns from more than 10,000 hours of mixed human video and motion data, then uses robot trajectories to adapt its predictions into whole body commands.[1][14]
- The model achieved a 71.1% aggregate success rate on the SIMPLE simulation benchmark, but its real robot tests were small and its results did not lead on every task.[6][5]
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is BeingBeyond’s Being M0.7, released by the embodied AI startup founded by Peking University tenured associate professor Lu Zongqing a. Article summary: Being M0.7 is BeingBeyond’s latent world action model for humanoid “loco manipulation”—coordinating walking and object manipulation—rather than treating them as separate skills.[1][6] Its central idea is to learn a predi. Topic tags: general web, ai, code, privacy, regulation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts w
Being-M0.7 is BeingBeyond’s model for humanoid “loco-manipulation”: moving around while reaching for or handling objects. The goal is to coordinate the whole body, instead of treating walking and manipulation as separate problems.1
6
The approach begins with more than 10,000 hours of raw human video and motion data. Only later does the model train on robot trajectories, which adapt what it has learned to a specific robot’s viewpoint, body and movement.1
14 BeingBeyond was founded by Peking University associate professor Lu Zongqing.
3
Three stages, from human data to robot commands
- Pretraining on human data: A Mixture of Transformers learns a shared video-and-motion representation and predicts future visual states alongside body motion. The aim is to capture patterns in how scenes and bodies change over time.
1
8
- Mid-training on robot trajectories: The model is adapted to the robot’s camera perspective, physical configuration and dynamics.
7
- Action-expert post-training: A separate action expert uses the frozen model’s predicted visual context to generate executable whole-body commands. This step grounds the human-derived prior in robot control; it does not simply replay human movement.
The training data includes paired egocentric video and motion, video-only samples, and motion-only samples. Paired examples provide joint video-and-motion supervision, while single-modality data broadens the training set.14 Reported sources include Ego4D, Xperience, Nymeria, Bones-SEED, SnapMoGen, HumanML3D and Lafan1, as well as earlier BeingBeyond datasets and commercial sources.
1
The model works with frozen DINO image features rather than predicting video pixels, and uses a compact, shared 22-dimensional, head-rooted motion representation.7
12 The available descriptions do not provide enough detail to explain how learnable placeholders for missing modalities are trained.
Why start with human data?
Robot demonstrations can be difficult to collect at scale. Human video and motion offer a much larger source of examples, but they are imperfect training material: some records have video without motion data, others have motion without video, and human movements cannot be transferred directly into a robot’s controls.5 Being-M0.7’s staged design aims to learn general patterns from this mixed data first, then use robot trajectories to adapt those patterns to a particular machine.
1
7
What the tests show
On SIMPLE, a simulation benchmark covering six tasks across three difficulty levels, Being-M0.7 recorded 128 successes in 180 trials, or 71.1%—the highest aggregate rate among the compared models.6 The reported comparison gives Ψ0 a 64.4% success rate, π0.5 52.8% and GR00T-N1.6 33.9%.
6 Without human-data pretraining, Being-M0.7’s score fell to 48.9%.
6
That overall lead did not translate into a win on every individual task: Ψ0 did better on one task, while π0.5 did better on two.6 The paper also reports that, compared with a language-context-only version, the full model improved on XMovePick and BendPick but did worse on XMoveBendPick. The authors attribute that result to possible interference between jointly trained tasks with similar visual observations.
5
In real-world tests on a Unitree G1 humanoid, the model was evaluated on three loco-manipulation tasks, with five trials per task. The reported result was a tie with the strongest baseline, not an outright lead.5 Because the evaluation involved only a small number of trials, it is not broad evidence of performance across different environments.
5
What remains to be tested
The evaluation covers a limited task horizon, so it does not establish how reliably the model handles longer sequences of actions.5 The researchers also reported that limited whole-body tracking accuracy from the SONIC controller made it harder to collect robot demonstrations. That constraint affects the quality and quantity of robot data available for training; it is a controller issue, not a demonstrated failure of the world-action model itself.
5
The October 12, 2026 date refers to coverage of the model; BeingBeyond’s research page lists Being-M0.7 on July 14, 2026.16
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Being-M0.7 uses human motion data to teach humanoid robots to move and manipulate"?
Being M0.7 is a latent world action model designed to coordinate humanoid walking and object manipulation, rather than training them as separate skills.[1][6]
What are the key points to validate first?
Being M0.7 is a latent world action model designed to coordinate humanoid walking and object manipulation, rather than training them as separate skills.[1][6] It learns from more than 10,000 hours of mixed human video and motion data, then uses robot trajectories to adapt its predictions into whole body commands.[1][14]
What should I do next in practice?
The model achieved a 71.1% aggregate success rate on the SIMPLE simulation benchmark, but its real robot tests were small and its results did not lead on every task.[6][5]
Sources
tech-momentum.comBeingBeyond, 사람 영상 1만 시간으로 휴머노이드 걷기·집기 학습
research.beingbeyond.comBeing-M0.7: A Latent World-Action Model for Humanoid Robots
ai.42how.com智在无界卢宗青团队提出 Being-M0.7，依托人类视频动作先验，训练人形机器人全身移动操作