A Learning-based Adaptive Compliance Method for Symmetric Bi-manual Manipulation
Learning-based Adaptive Compliance Method (LAC)
We propose a novel Learning-based Adaptive Compliance (LAC) algorithm to improve the efficiency and adaptability of symmetric bi-manual manipulation.
Videos
The video intuitively shows the effectiveness of LAC in the dual-arm cooperative handling and the peg-in-hole assembly operations.
Overview
We further made the following contributions:
We design a novel algorithm framework to complete symmetric bi-manual manipulation based on reinforcement learning. The LAC generates the desired trajectory of dual-arm operations, improving motion planning efficiency. Meanwhile, it adaptively updates impedance parameters to overcome the poor robustness of impedance control.
We design a centralized Actor-Critic network structure with LSTM networks to achieve cooperative planning and impedance-parameter modification. The centralization of the framework is particularly useful in operations requiring high dual-arm synchronization. Moreover, results demonstrate that the force trend feature captured by LSTM networks significantly impacts the performance of compliance operations.
Thanks to the efficiency and robustness of LAC, we realize two typical symmetric bi-manual manipulations, the dual-arm cooperative handling operation with random target positions and the peg-in-hole assembly operation in which two arms hold the same peg.
Method
We present the algorithm framework of LAC used in the symmetric bi-manual manipulation. It is a centralized framework consists of two parts. The high-level module is the reinforcement learning running at 20 Hz that provided the object's desired trajectory and the parameters of the impedance controller. The low-level module is the impedance control running at a frequency lower than that of the high-level module.Â
Simulation
Environment
The simulation environments of the dual-arm cooperative handling and peg-in-hole assembly are built in Mujoco.
Comparision
The following figures and tables compare the performance of different algorithms in the two practical operations.