Constraint-Manifold-Guided Diffusion and Mode-Aware Compliance for Bimanual Assembly
Bimanual robotic assembly requires a deep synergy between global kinematic coordination and local physical stability, which is often hindered by complex coupling constraints and contact uncertainties. This work presents a hierarchical dual-manifold framework that decouples long-horizon generative planning from reactive contact adaptation. The planning layer employs a constraint-guided diffusion policy to ensure that generated trajectories conform to a geometric consistency manifold by incorporating task-level kinematic relationships as a differentiable guidance field. In synergy with this, a mode-aware residual policy operates on a physical interaction manifold, utilizing sparse kinesthetic corrections to provide fine-grained pose and wrench modulations for internal stress regulation. Both modules function within a receding-horizon architecture and are trained directly on real dual-arm hardware, bypassing the fidelity gaps inherent in simulated contact dynamics. We evaluate the system on a dual-UR5 platform across three representative tasks: rigid closed-chain insertion, semi-constrained cable routing, and decoupled stabilize-and-act insertion. Experimental results demonstrate that the framework consistently outperforms classical, learning-based, and ablated baselines across three coordination regimes, achieving 88–92% success rates while reducing interaction loads and improving contact robustness during bimanual assembly. Note to Practitioners—The motivation of this study is to address the challenge of jointly handling long-horizon geometric coordination and local contact adaptation in industrial bimanual assembly. Existing approaches often struggle to balance global motion consistency with rapid correction of contact-induced deviations under dual-arm coupling. We propose a hierarchical dual-manifold framework: a constraint-guided diffusion policy generates globally consistent dual-arm references, while a mode-aware residual policy compensates for local deviations during execution. Trained directly on real robot data, the framework reduces reliance on high-fidelity contact simulation and demonstrates robust performance across assembly tasks with different coordination modes. Its current limitation is that it mainly focuses on the assembly execution stage and does not yet include preceding operations such as grasping. Future work will extend the framework to more complex industrial geometries and multi-stage assembly scenarios.