Learning Action Priors for Cross-embodiment Robot Manipulation

Researchers propose a two-stage training framework for Vision-Language-Action (VLA) models that pretrains the action module with motion priors before multimodal alignment. This approach enables robots to learn temporal dynamics more efficiently and generalizes better across different embodiments and real-world tasks with limited data.

This research addresses a fundamental limitation in current Vision-Language-Action models: the action module must learn physical motion dynamics while simultaneously aligning with visual and linguistic features, creating a bottleneck in policy optimization. The proposed solution decouples these challenges through staged training, where the action module first absorbs motion structure from raw trajectory data using flow-matching techniques, then transfers this learned prior to VLA training. This architectural insight reflects broader progress in machine learning toward modular, transfer-friendly designs.

The approach builds on established principles in robotics and deep learning—that task-specific priors accelerate convergence and improve generalization. However, applying this principle to cross-embodiment settings (enabling one trained model to control different robot morphologies) is non-trivial and represents meaningful technical progress. The framework's use of a compact history compressor derived from the pretrained encoder also demonstrates efficiency gains relevant for real-world deployment constraints.

The experimental validation across 13 tasks spanning simulation and physical robots shows the practical value of the method, particularly in data-scarce scenarios common in real-world robotics. Faster convergence and higher success rates directly reduce development costs and enable broader adoption of robot learning systems. The finding that scaling action pretraining data improves downstream performance suggests a new frontier for robot learning: leveraging large unlabeled trajectory datasets independently of specific downstream tasks.

• →Two-stage training with motion priors reduces simultaneous optimization challenges in cross-embodiment robot learning
• →Pretrained action modules improve performance significantly on real-world tasks with limited labeled data
• →The approach enables compact state-action history compression at minimal computational cost
• →Scaling motion pretraining data yields generalizable priors that transfer to diverse downstream VLA tasks
• →Cross-embodiment generalization improves substantially compared to end-to-end VLA training without action priors