Search papers, labs, and topics across Lattice.
University of Michigan, Ann Arbor
3
0
4
Polynomial representations in robot control can outperform larger MLPs, revealing that interaction structures are crucial for effective motor policies.
GaitSpan achieves a continuous range of locomotion from walking to running with a single adaptable policy, outperforming traditional multi-expert and imitation-based approaches.
Humanoids can now recover from falls in complex environments without real-world training, thanks to a distilled, goal-in-context policy that reasons about both pose and terrain.
