Legged locomotion through deep mud — a deformable, flow-prone medium where the usual assumptions about ground contact break down.
Mud is deceptively hard to walk on: it behaves like a solid under light loads but flows like a fluid under heavier ones, so a robot's feet sink, get held by suction, and lose traction unpredictably. There is no single "right" gait, and the rigid-contact models most locomotion controllers depend on simply don't hold.
This project trains a model-free, learned controller — with no explicit contact or terrain model at run time — to keep a quadruped stable and efficient across varying, deformable soil. In a terramechanics-based simulation of how mud yields and flows underfoot, the policy learns to adapt its gait as conditions change.
If you're interested in this project or a potential collaboration, feel free to reach out.