BSD-PPO: A Fusion Algorithm for Dual-Arm Robot Collaborative Optimization and Dynamic Collision Avoidance

To address the issues of inter-arm collisions, difficulty in coordinating multimodal actions, and the lack of real-time optimization in offline policies for bimanual robot manipulation, this paper proposes the Bimanual Synergistic Diffusion-PPO (BSD-PPO) framework. This framework combines the multimodal generation capability of diffusion models with the online optimization mechanism of PPO, constructing a DDPM-based “pre-generation–online optimization” two-layer architecture. Actions are proposed via the Diffusion Policy and optimized through the Actor-Critic mechanism, while cooperative advantage estimation, Clipped Objective, VLM collision-avoidance instructions, and force-feedback adaptive decoding are introduced to enhance collision detection and computational efficiency. The framework includes key modules such as the Diffusion Policy and PPO optimizer, with the operational process divided into four stages, providing an effective solution for embodied intelligence bimanual manipulation.