jackzhousz / opt-mimic-traj-opt

Trajectory optimization code for the paper "OPT-Mimic: Imitation of Optimized Trajectories for Dynamic Quadruped Behaviors"

opt-mimic-traj-opt

Trajectory Optimization code for the paper "OPT-Mimic: Imitation of Optimized Trajectories for Dynamic Quadruped Behaviors".

Quick Reference

• traj_opt.py contains the core implementation of the trajectory optimization, including some constraints that need to be modified specifically to each motion.
• generate_reference.py contains code producing hard-coded kinematic sketch trajectories via sinusoids and interpolation of keypoints. The motion type is specified using the motion_type variable (line 41 as of writing).
• python main.py -d 0 -s 1 -e 1 -n my-experiment-name is a typical command I would run to run trajectory optimization, and save video and csv results under filename my-experiment-name. You likely need to create an empty videos/ folder to avoid getting an error the first time.
• constants.py contains parameter values, of which I occasionally had to tune LQR weights specifically to each motion.

Integration with Other Code

• Once a csv trajectory file is saved onto csv/, it should be copied onto the raisimGymTorch/raisimGymTorch/env/envs/solo8_env/traj/ directory of the reinforcement learning code to use it to train RL policies.
• When deploying the trained RL policy on the robot, the csv trajectory file also needs to be copied onto the workspace directory containing the robot deployment code. See that repo for details.