← current work

Current work

Learned Locomotion Through Underbrush

Dense vegetation and vine-like entanglements are some of the hardest terrain for a mobile robot to cross.

This project develops proprioceptive sensing and reactive control that let quadrupedal robots feel, respond to, and push through tangled vegetation instead of getting stuck. Perception alone is unreliable here — foliage occludes the scene and depth sensors routinely misjudge the true height of the ground plane — so we lean on leg force feedback, paired with learned disentanglement behaviors, to traverse cluttered natural terrain robustly.

Stock Unitree controller: the legs snag and tangle in compliant vines, and forward progress stalls.
Learned disentanglement behavior: the robot senses the snag, works its legs free, and keeps walking.