BeingBeyond / Being-H0.7
- latent actions
- 2026
- World action model
Model cardPaper ↗
Being-H0.7 predicts future interaction in a compact latent space, then acts without generating future images at test time.
Key facts
Rows below are the register record in lib/models.ts. A dash means the value is not confirmed from a primary source. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and the model's own paper.
| Name | Being-H0.7 |
|---|---|
| Lab | BeingBeyond |
| Year | 2026 |
| Paradigm | latent actions |
| Backbone | — |
| Source | https://research.beingbeyond.com/being-h07 |
What the paper claims
- The model is trained on 200,000 hours of egocentric human video and 15,000 hours of robot demonstrations. A posterior branch sees future observations during training; at test time only the prior branch runs, using learnable latent queries as the world-action workspace. BeingBeyond, 14 April 2026
- Reuss describes the same prior/posterior split as Play-LMP scaled into a Mixture-of-Transformers, with InternVL3.5 as the understanding expert, Qwen3 as the action expert, frozen V-JEPA 2.1 future embeddings, and a flow-matching action policy. Those component names are Reuss's reading of the architecture; the project page states the latent prior/posterior idea and the data mix without repeating every backbone name. Reuss, NVIDIA Technical Blog (15 June 2026)
What it evaluates
- On six simulation benchmarks the authors report the strongest overall performance and the best average rank in their comparison, covering LIBERO, LIBERO-plus, GR1, both CALVIN splits, RoboCasa and Robotwin2. BeingBeyond, 14 April 2026
- Real-world evaluation uses 12 tasks, three platforms (PND Adam-U, Unitree G1, Franka FR3) and five suites: Dynamic Scene, Physical Reasoning, Motion Reasoning, Long Horizon and Generalization. They report leading all five suites against Being-H0.5, π0.5 and Fast-WAM. BeingBeyond, 14 April 2026
- With Universal Async Chunking they report exposed step latency for Being-H variants in the 3–4 ms range. On Unitree G1 they pair the upper-body policy with a pretrained AMO whole-body backend. BeingBeyond, 14 April 2026
On this register
Latent actions. Learns actions as hidden codes from video. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other latent actions models on the register: DreamDojo.
The backbone field is unconfirmed, so this card does not claim a video-model family.