Abstract
Bimanual manipulation has been widely applied in household services andmanufacturing, which enables the complex task completion with coordinationrequirements. Recent diffusion-based policy learning approaches have achievedpromising performance in modeling action distributions for bimanualmanipulation. However, they ignored the physical safety constraints of bimanualmanipulation, which leads to the dangerous behaviors with damage to robots andobjects. To this end, we propose a test-time trajectory optimization frameworknamed SafeBimanual for any pre-trained diffusion-based bimanual manipulationpolicies, which imposes the safety constraints on bimanual actions to avoiddangerous robot behaviors with improved success rate. Specifically, we designdiverse cost functions for safety constraints in different dual-arm cooperationpatterns including avoidance of tearing objects and collision between arms andobjects, which optimizes the manipulator trajectories with guided sampling ofdiffusion denoising process. Moreover, we employ a vision-language model (VLM)to schedule the cost functions by specifying keypoints and correspondingpairwise relationship, so that the optimal safety constraint is dynamicallygenerated in the entire bimanual manipulation process. SafeBimanualdemonstrates superiority on 8 simulated tasks in RoboTwin with a 13.7% increasein success rate and a 18.8% reduction in unsafe interactions overstate-of-the-art diffusion-based methods. Extensive experiments on 4 real-worldtasks further verify its practical value by improving the success rate by32.5%.