Embedded Files
Abstract
Abstract
This project aims to reproduce the results presented in the paper titled Learning Agile Robotic Locomotion Skills by Imitating Animals on our self-designed multi-modal quadruped robot. I was involved in this project when I was doing my research internship at Tencent Robotics X Lab and my work mainly focus on the Sim2Real part. During my internship, I managed to transfer gaits learned in simulation to the real quadruped robot with a 100% success rate.
Pipeline
Pipeline
sim2real_mocap.mp4The main goal of this project is to reproduce pipeline in the left image to our custom quadrupled robot as shown in the right video.
Key Factors
Key Factors
real_world_mocap_walk_crash_down2.mp4Performance at the beginning of my internship
Methods conducted enabling successful Sim2Real transfer
Methods conducted enabling successful Sim2Real transfer
Hardware calibration: IMU and joint configuration state; URDF information confirmation
Simpler tasks: trajectories from MPC with controlled maximal velocity;
Dog walking skill learning with realistic styles: residual reinforcement learning, longer historical information.
trot_walk_history_act.mp4Performance at the end of my internship
Key factors found for successful Sim2Real during my internship
Key factors found for successful Sim2Real during my internship
Hardware calibration always before transferring you policy;
Domain Randomization on joint configurations can handle sensor mismatch or small calibration error;
Use longer history of state and action information to handle model mismatch
Page updated
Google Sites
Report abuse