LingBot-Video opens a video world model for robot actions
LingBot-Video predicts future robot movement as video from action and hand-pose inputs. The model has 13 billion in total, but only 1.4 billion are active during each run because it uses a design. This aims to keep the capacity of a large model while lowering the amount of computation used at once.
After training, it was refined with six rewards, including a reward for physical plausibility. One concern is that a VLM judges that physical plausibility from sampled video frames, which may reward surface-level tricks instead of real physics understanding. The model is described as useful for policy and action planning, but the shown results are mainly video-quality results, not closed-loop robot numbers.
It has the best average score on RBench, while a still leads on the more reasoning-heavy parts, and it ranks second on general text-to-video in its own . The weights, code, and Diffusers/SGLang stack are publicly available.
Key points
- The model has 13B total but uses 1.4B per run.
- It can predict robot rollouts from action and hand-pose conditions.
- It was refined with six rewards, including physical plausibility.
- A VLM judges physical plausibility, which raises concerns.
- No closed-loop robot numbers are provided yet.