CoFreeVLA: Collision-Free Dual-Arm Manipulation via Vision-Language-Action Model and Risk Estimation

TL;DR

CoFreeVLA tackles the unsafe self-collisions in dual-arm Vision-Language-Action manipulation by integrating a short-horizon self-collision risk estimator into execution, recovery, and policy refinement. The estimator, trained with model-based labels and real-robot calibration, predicts $\hat{r}_t$, $\hat{d}_{\min}$, and $\hat{\tau}_{\text{ttc}}$ from $s_t$ and $A_{t:t+H-1}$ and gates actions using thresholds $\tau_\uparrow$ and $\tau_\downarrow$. Across five real dual-arm tasks, CoFreeVLA reduces self-collisions and maintains task performance relative to strong baselines like RDT and APEX, with sub-5 ms inference enabling real-time operation. The framework demonstrates that incorporating explicit, short-horizon risk reasoning into a VLA pipeline yields safer, more reliable dual-arm manipulation in practical settings. The approach blends model-based pretraining, real-world calibration, and risk-aware policy refinement to deliver practical improvements for collision-free dual-arm robotics.

Abstract

Vision Language Action (VLA) models enable instruction following manipulation, yet dualarm deployment remains unsafe due to under modeled selfcollisions between arms and grasped objects. We introduce CoFreeVLA, which augments an endtoend VLA with a short horizon selfcollision risk estimator that predicts collision likelihood from proprioception, visual embeddings, and planned actions. The estimator gates risky commands, recovers to safe states via risk-guided adjustments, and shapes policy refinement for safer rollouts. It is pre-trained with model-based collision labels and posttrained on real robot rollouts for calibration. On five bimanual tasks with the PiPER robot arm, CoFreeVLA reduces selfcollisions and improves success rates versus RDT and APEX.

Figures (2)

• Figure 1: (a) Risk Estimator. Given a state–action pair, the model applies a cross-attention module followed by three heads to predict collision risk $r$, minimum distance $d$, and time-to-collision $\tau$. (b) Data Collection and (c) Dataset. For each state, we sample candidate actions and check their outcomes for labeling. (d) CoFree VLA. At run time, the system continuously estimates risk; when it exceeds a threshold, the policy will be blocked and the arms will be returned to a safe configuration, after which the policy resumes to complete the task.
• Figure 2: Experimental results in 5 dual-arm manipulation tasks. Evaluation of collision and success rates is on 10 trials.