Vision Language Action (VLA) models enable instruction-following manipulation, yet their deployment on coordinated dual-arm platforms remains severely constrained by under-modeled self-collisions between manipulators and grasped objects. To address this critical safety gap, we propose CoFreeVLA, a novel framework that augments end-to-end VLA policies with a lightweight, short-horizon self-collision risk estimator. The estimator predicts collision likelihoods directly from proprioceptive states, visual embeddings, and candidate action sequences. Deeply integrated into the closed-loop control system, this estimator proactively gates risky commands, autonomously synthesizes recovery trajectories to safe states via risk-guided adjustments, and biases policy refinement for safer rollouts. To ensure robust calibration, the estimator utilizes a two-stage training pipeline, pre-training with model-based synthetic collision labels, followed by post-training on real-robot rollouts. Across five bimanual tasks, six VLA backbones, and 30 trials per variant, the task-averaged collision rate decreases from 0.54 to 0.23, while the task-averaged success rate increases from 0.45 to 0.61. Compared to representative baselines, CoFreeVLA substantially reduces self-collision frequencies and improves overall task success rates, providing a crucial step toward the safe deployment of foundational models in multi-arm continuous control.

翻译：暂无翻译