AnswersPublished19 sources
What is BeingBeyond’s Being-M0.7, released by the embodied AI startup founded by Peking University tenured associate professor Lu Zongqing and described in an October 12, 2026 post, and how does it enable simultaneous humanoid walking and object manipulation by using more than 10
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 predictive video–motion prior from more than 10,000 hours of raw human data before us Being-M0.7 is Be...
Δημοσιεύτηκε απόΕπεξεργασία με GPT-6 LunaΕικόνες δημιουργήθηκαν με GPT Image 2
Research answer
Key takeaways
- 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 predictive video–motion prior from more than 10,000 hours of raw human data before us
- 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 predictive video–motion prior from more than 10,000 hours of raw
- ## How it works and why human data helps
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 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 predictive video–motion prior from more than 10,000 hours of raw human data before using robot trajectories to adapt that knowledge into executable whole-body commands.
1
8
How it works and why human data helps
- Motivation: Human data can supply much broader experience than robot demonstrations alone, but its modalities are inconsistent: some samples contain only video, some only motion, and human movements cannot be transferred directly into robot commands.
5 The approach therefore separates learning how scenes and bodies evolve from learning how a particular robot must act.
1
7
- Pretraining: A Mixture of Transformers learns a shared video–motion space and jointly predicts future visual representations and body motion, producing a human-centric predictive prior.
8
- Mid-training: Robot trajectories adapt that prior to the robot’s camera viewpoint, body configuration, and dynamics.
- Post-training: A separate, gated action expert learns to turn the frozen prior’s predicted visual context into executable whole-body commands.
The earlier search results identify three complementary training streams—paired egocentric video and motion, video-only data, and motion-only data—with paired examples supplying joint supervision and unpaired examples expanding coverage.5
9 They also identify Ego4D, Xperience, Nymeria, Bones-SEED, SnapMoGen, HumanML3D, and Lafan1, supplemented by subsets of earlier Being-H0.5 and Being-M0.5 corpora and commercial sources.
1
Representation details: evidence boundary
- Earlier retrieved material describes DINO visual latents and a compact, shared 22-dimensional head-rooted motion interface; it also says action-expert training reads future visual hidden states while explicit motion prediction is disabled.
14
- However, the supplied excerpts do not establish the precise DINO-freezing procedure, the 22 coordinates’ definitions, or how learnable missing-modality placeholders are trained. Insufficient evidence to explain those implementation details confidently.
- Likewise, the excerpts establish the motivation for scaling with human data, but do not quantify robot-demonstration scarcity or collection cost.
Results and task-level qualifications
- SIMPLE: The available evidence confirms six tasks at three difficulty levels and 128 successes in 180 trials, or 71.1%, the highest aggregate success rate among the compared models.
- Not every task improves: Relative to a language-context-only variant, the full model improves XMovePick and BendPick but performs worse on XMoveBendPick; the reported explanation is interference between jointly trained tasks with similar visual observations.
- Real Unitree G1: The evidence confirms three real-world loco-manipulation tasks, five trials per task, and a tie with the strongest baseline—not an outright lead.
The question supplies additional figures: SIMPLE baselines of Ψ0 64.4%, π0.5 52.8%, and GR00T-N1.6 33.9%; a no-human-pretraining ablation of 48.9%; and G1 totals of 13/15 for both Being-M0.7 and GR00T-N1.6, versus 7/15 for π0.5 and 3/15 for Ψ0. These are consistent with the aggregate ranking described above, but the returned excerpts do not independently expose those full numerical tables. Insufficient evidence to verify all those numbers or list every individual-task baseline exception.
Limitations and interpretation
- Small real-world evaluation: Only five trials per task are documented, so the reported tie should not be read as broad evidence of equal reliability across environments.
- Short horizons: Short walk-and-grasp tests would not establish dependable long-duration operation, multi-step task completion, or recovery from accumulating errors. The excerpts do not provide enough detail to verify the paper’s exact horizon limitation.
- Demonstration controller: Inaccurate whole-body tracking during demonstration collection could introduce a mismatch between intended motion and recorded execution, limiting supervision quality. This is a relevant methodological concern, but the supplied evidence does not establish the controller’s measured accuracy or the authors’ exact claim.
- Do not import another model’s limitation: The supplied open-loop/replanning limitation belongs to ZeroWBC, not Being-M0.7.
2
Finally, October 12 is the date of the referenced coverage, not necessarily the original release: an earlier retrieved source reports a July 2026 release.4 The evidence identifies Lu Zongqing as a Peking University associate professor and BeingBeyond’s founder, but does not independently verify the “tenured” qualifier.
6
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 "What is BeingBeyond’s Being-M0.7, released by the embodied AI startup founded by Peking University tenured associate professor Lu Zongqing and described in an October 12, 2026 post, and how does it enable simultaneous humanoid walking and object manipulation by using more than 10"?
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 predictive video–motion prior from more than 10,000 hours of raw human data before us
What are the key points to validate first?
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 predictive video–motion prior from more than 10,000 hours of raw human data before us 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 predictive video–motion prior from more than 10,000 hours of raw
What should I do next in practice?
## How it works and why human data helps