We demonstrate a working implementation of Policies Modulating Trajectory Generators (Iscen et al. 2019) on the MIT Mini Cheetah. When the trajectory generator is tuned to accommodate more dynamic behaviors, unrealistic and non-robust motion artifacts emerge in simulation due to the implicit treatment of force.
Successful PMTG Deployment (Flat Terrain)
TrotFwd.mp4Forward Trot
PMTGDeployFlat1.mp4Testing robustness against external push
PMTG Simulation Results (Gap Terrain)
Relaxed TG Parameters enable longer gap crossing, but introduce undesired gait characteristics such as missed contacts and rapid changes in roll and pitch.
Pronking
Pronk15cmConservative.mp4Conservative TG parameters
Max Gap Width : 15cm
Pronk25cmSparse.mp4Relaxed TG parameters
Max Gap Width : 25cm
Trotting
Trot10cmConservative.mp4Conservative TG parameters
Max Gap Width : 10cm
Trot15cmSparse.mp4Relaxed TG parameters
Max Gap Width : 15cm
Trajectory Plots
PMTG without Specialized Rewards (Gap Terrain)
We found that PMTG baseline requires specialized reward design and extensive tuning for traversing gaps larger than 10cm in case of Trotting and 20cm in case of Pronking. Otherwise, it leads to unsuccessful gap crossing behaviors such as gap avoidance or in-place motion.
TrotNotTuned.mp4Gap Width : 10cm
PronkNotTunned.mp4Gap Width : 20cm