Grasp, Handover, Rotate: Bimanual Object Reorientation via Compositional Diffusion and Energy-Based Optimization

Researchers at The Hong Kong University of Science and Technology and Shenzhen Loop Area Institute introduce BiCompoDiff, a unified framework for bimanual object reorientation: one UR12e arm grasps and hands over an object, while the other regrasps and places it in a new pose. BiCompoDiff combines the pretrained GraspGen 6-DoF diffusion model with differentiable energy-based guidance for collision avoidance, handover feasibility, regrasp safety, and joint-space smoothness. During reverse diffusion, annealed MCMC refines grasp, handover, and placement poses, while the learned SubnetIK model supplies fast gradients for inverse-kinematics feasibility; cuRobo then generates minimum-jerk trajectories. On a benchmark of 60 simulated tasks spanning easy, medium, and hard clutter, the full system reduced joint displacement by 37% and achieved a 63.3% success rate versus 41.7% without planning-energy guidance. Against an adapted ReorientBot baseline, it reached 81.7% success versus 58.3%, a 23.4% absolute improvement, while also reducing joint displacement by 12.5%. Real-robot tests transferred the method to two UR12e arms with Robotiq grippers: joint displacement improved by 59.7% for a T-shaped brick and 46.4% for a cup. The results demonstrate that compositional diffusion can optimize grasp quality and downstream motion jointly rather than relying on costly sample-and-filter pipelines. The project involves HKUST and Shenzhen Loop Area Institute and acknowledges support from the BYD-HKUST Joint Lab.