Appendix B
Glossary
about 1 minutes
- Action chunk. The 32×7 block of future actions the model emits per query.
- Macro-step. One policy query and the 10 environment steps that execute part of its chunk; the RL time step.
- Flow matching. A generative model that learns a velocity field transporting noise to data; sampling integrates that field.
- Flow-SDE. The noise-injected version of the flow sampler with the same marginals and Gaussian per-step transitions, enabling log-probabilities.
- MoT. Mixture of transformers; here, the video and action experts running side by side with cross-attention.
- K/V cache. The per-layer keys and values of the first-frame video tokens that action tokens attend to; the "features the actor depends on".
- GRPO / group-relative advantage. Normalising returns within a group of rollouts from the same start instead of using a learned value function.
- PPO clip. Limiting the probability ratio between new and old policy to in the objective.
- Video SFT. Continuing the world model's supervised video-prediction training on new data.
- LoRA. Low-rank adapters added to a frozen network so that only a small number of parameters are trained.