B.Sc. thesis · Amirkabir University
Real-time MPC for a quadruped robot
A quadruped that keeps walking when it is handed a payload it was not told about.
The problem
Model Predictive Control is only as good as its model. When a legged robot picks up an unknown load, the model is suddenly wrong: the robot slows down, tilts and tracks its commands poorly.
What I did
- Full-body Newton–Euler dynamic model of the quadruped
- Real-time MPC in MuJoCo for walking and running
- An online payload identifier that estimates the added mass and center-of-mass shift and updates the MPC model
Results
In each video the same robot gets a payload partway through, walking or running. With the nominal model (left) it falls behind the commanded speed, and with 10 kg on its back it falls over. With the payload identifier (right) it estimates the new load within seconds and keeps tracking the command upright.