An attention-based encoder and recurrent memory process raw lidar returns. A phase-scheduled teacher-student pipeline combines separate jumping, brachiation, and landing experts into a perceptive policy. Hardware transfer includes sensor noise and actuator constraints, while passive hook end effectors support robust bar contact.
The work is useful for studying agile whole-body control under sparse perception. The reported hardware evaluation completes 14 of 15 traversals across three bar configurations and also demonstrates a separately trained ducking policy. These are research results on a specific setup, not a downloadable general-purpose robot controller.
