Summary
ThinkingVLA is a VLA model that decomposes manipulation planning into forward and inverse chain-of-thought reasoning within a unified Mixture-of-Transformers architecture. A forward CoT identifies the immediate subgoal and guides visual forecasting of the predicted next state; an inverse CoT then reasons about spatial relationships and action intent conditioned on that predicted image before generating the final action.
Key Contributions
- Unified Mixture-of-Transformers (MoT) architecture jointly handling visual prediction and action generation
- Forward CoT identifies subgoals and predicts the target visual state as an intermediate representation
- Inverse CoT grounds action generation in the visually predicted future state, enabling richer spatial reasoning
- Consistent outperformance over state-of-the-art baselines, especially on long-horizon manipulation tasks
Significance
Addresses the key limitation that standard VLAs map observations directly to actions without intermediate reasoning, making them brittle on long-horizon tasks; ThinkingVLA’s interleaved vision-language reasoning is a strong step toward interpretable, compositional robot policies.