Summary
Current VLA models leave the action module to learn physical motion from scratch during joint training, requiring simultaneous discovery of temporal action dynamics and cross-modal alignment—a challenge amplified in cross-embodiment settings. This work pretains the action module with motion priors before cross-modal VLA alignment via a two-stage framework: Stage 1 learns temporal motion structure from action trajectories alone using a flow-matching encoder-decoder (no visual/language tokens); Stage 2 transfers the prior to VLA training via decoder reuse and early-stage latent distillation.
Key Contributions
- Decoupled two-stage training: motion prior pretraining then cross-modal VLA alignment
- Stage 1: Lightweight flow-matching encoder-decoder learns temporal structure from unconditioned action trajectories
- Stage 2: Learned encoder used as compact history compressor, summarizing state-action histories into a single temporal context token
- Designed for cross-embodiment generalization with heterogeneous action distributions
Significance
Giving the VLA action module an explicit motion prior before cross-modal training reduces the simultaneous optimization burden and yields better cross-embodiment transfer with less data.