Learned Swing Leg Trajectory Optimization

This project has code on Github here.

Along with a WandB report here.

We use pydrake to create an offline nonlinear optimization pipeline using direct collocation to solve for obstacle-aware, dynamically feasible legged trajectories. While the output paths are close to optimal, the solve time for the pipeline is on the order of 1-4 minutes, making it impractical for realtime gaits.

We run some experiments trying to train small networks to output the same trajectory using an input heightmap. These small networks can be run in inference mode on embedded hardware at 50-100 Hz, making them much more practical for real-time use.