A Learning-based Adaptive Compliance Method for Symmetric Bi-manual Manipulation

TL;DR

This work tackles the inefficiency of separating motion planning from impedance control in symmetric bi-manual manipulation for on-orbit assembly by introducing Learning-based Adaptive Compliance (LAC). LAC uses a centralized actor-critic architecture with LSTM-based force-state preprocessing to jointly generate the object trajectory and adapt impedance parameters, delivering $X_d^o = X_g + a_x$ with $a_x \in \mathbb{R}^3$ and impedance updates $a_p \in \mathbb{R}^6$. The system operates a slow high-level policy at 20 Hz while a fast low-level impedance loop maintains stable control, and the approach is validated through both Mujoco simulations and real-world dual-arm peg-in-hole experiments showing improved robustness, generalization, and synchronization. Ablation and generalization studies, along with real-world tests on UR arms, demonstrate that LAC achieves high-precision assembly with manageable interaction forces, confirming its potential for robust on-orbit cooperative manipulation. The results indicate that integrating planning and adaptive compliance within a centralized RL framework can significantly enhance efficiency and reliability in complex dual-arm tasks.

Abstract

Symmetric bi-manual manipulation is an essential skill in on-orbit operations due to its potent load capacity. Previous works have applied compliant control to maintain the stability of manipulations. However, traditional methods have viewed motion planning and compliant control as two separate modules, which can lead to conflicts with the simultaneous change of the desired trajectory and impedance parameters in the presence of external forces and disturbances. Additionally, the joint usage of these two modules requires experts to manually adjust parameters. To achieve high efficiency while enhancing adaptability, we propose a novel Learning-based Adaptive Compliance algorithm (LAC) that improves the efficiency and robustness of symmetric bi-manual manipulation. Specifically, the algorithm framework integrates desired trajectory generation and impedance-parameter adjustment under a unified framework to mitigate contradictions and improve efficiency. Second, we introduce a centralized Actor-Critic framework with LSTM networks preprocessing the force states, enhancing the synchronization of bi-manual manipulation. When evaluated in dual-arm peg-in-hole assembly experiments, our method outperforms baseline algorithms in terms of optimality and robustness.

Figures (10)

• Figure 1: Closed-chain model of the dual-arm robot.
• Figure 2: Algorithm framework of LAC. (a) Workflow of symmetric bi-manual manipulation consists of three parts. (b) Adaptive compliance algorithm based on reinforcement learning.
• Figure 3: The two stages of the dual-arm cooperative peg-in-hole assembly task. The object needs to be first transported to the target position and then slowly inserted.
• Figure 4: The ground-based dual-arm robot system for typical on-orbit assembly operation.
• Figure 5: Block diagram of experimental system of adaptive compliance control algorithm based on reinforcement learning.