A robot policy can recognize what it sees and still fail to anticipate what its actions will do. PDRVLA addresses this gap by teaching the policy to represent future physical changes, while avoiding the latency of generating a future video at inference time.
Key Findings
- A lightweight Physical Dynamics Module extracts physical tokens from vision-language features and uses them to condition action generation. This makes predicted dynamics available to the policy without putting video synthesis in the control loop.
- The module predicts future RGB, depth, and optical-flow features, with auxiliary supervision from a frozen video world model. The design targets task-relevant changes rather than requiring online generation of complete future scenes.
- The paper reports success rates of 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin 2.0, and 81.5% on SimplerEnv. These results are simulation benchmarks, not evidence of equivalent performance on physical robots.
- Training uses publicly available data augmented with human-to-robot synthesis, subtask annotations, and embodied chain-of-thought annotations. Ablations examine how these signals affect future prediction and task success.
Who Should Read It
A good fit for researchers studying predictive representations and latency-aware vision-language-action policies, especially those comparing explicit dynamics supervision with reactive action prediction. Look elsewhere if you need demonstrated real-world deployment results or a direct comparison of inference latency against other methods; the reported evidence here centers on simulation performance and ablations.