EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards
TLDR
EVA uses inverse dynamics rewards to align video world models with executable robot actions, addressing the executability gap via RL post-training.
Reasoning
The paper clearly identifies a practical problem (executability gap) and proposes a novel RL-based alignment method using inverse dynamics as reward. Strengths include direct use of real robot trajectories and robustness to visual artifacts. Weaknesses include potential embodiment specificity and reliance on pre-collected real data.
Read-first score
Read-first score 71.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.
Field roles
Rank sensitivity
Stability: volatile; rank range: 85.
Keyword Scores
Deep Analysis
Innovations
- Leveraging the executability gap as a training signal for video world models
- Repurposing an inverse dynamics model (IDM) trained on real robot trajectories as a reward model for video alignment
- Reinforcement-learning post-training framework (EVA) to align video world models with executable robot actions
Methodology
EVA is a reinforcement-learning post-training framework that trains an inverse dynamics model (IDM) on real robot trajectories and uses it as a reward model to evaluate generated videos through the action sequences they induce. The reward encourages smooth motions (measured by velocity, acceleration, and jerk) and penalizes actions that violate embodiment constraints, even when generated videos contain severe visual artifacts.
Key Results
Experiments on the RoboTwin benchmark and a real bimanual robot show that EVA reduces embodiment-specific artifacts in generated rollouts and improves downstream task execution success.