Multi-Modality Driven Impedance-Based Sim2Real Transfer Learning for Robotic Multiple Peg-in-Hole Assembly
- Quick Demos
video_assembly.mp4
- Run
conda create -n dynamic-assembly python=3.8.13 && conda activate dynamic-assembly
to create and activate a new python environment. - Install MuJoCo using these instructions (i.e. extract the downloaded
mujoco210
directory into ~/.mujoco/mujoco210) - Use obj2mjcf to process original robot urdf file and assembly-related object file into XML file for the use in MuJoCo.
- Run
cd dynamic-assembly && ./install-dependencies.sh
to install all required dependencies.
- Collect a dataset from different hole/peg shapes during raw policy learning based on the ShuffleNet-v2 perception. An example dataset of circle shape can be obtained
https://drive.google.com/drive/folders/181-17Ub87fH-swqxKHRSpU4iK7TywIp7?usp=drive_link
. - The pretarining framework is designed from simple VAE models. You can test the collected dataset based on these models.
- Modified from the robosuite, we use our own Kuka_iiwa robot and self-designed peg/hole objects to execute the assembly task. Model files are in
robosuite/models
and the assembly enviroment settings inrobosuite/enviroments
. - The manipulation of robot model and robot controllers are based on the MujocoPy.
- Install the package of Viskit from rllab.