Learning to Jump from PixelsDownload PDF

Jun 19, 2021 (edited Aug 25, 2021)CoRL2021 PosterReaders: Everyone
  • Keywords: Locomotion, Vision, Hierarchical Control
  • Abstract: Today's robotic quadruped systems can robustly walk over a diverse range of natural but continuous terrains involving snow, rain, slip, rubble, etc. Locomotion on discontinuous terrains such as one with gaps or obstacles presents a complementary set of challenges. It becomes necessary to plan ahead using visual inputs and execute agile behaviors such as jumps to cross gaps. Such dynamic motion results in significant motion of on-board camera that introduces a new set of challenges for real-time visual processing. The need for agility and the operation from vision reinforce the need for robust control. We present a system that can, in real-time, process visual observations from an onboard RGBD camera to command a quadruped robot to jump over wide gaps. The proposed method brings together the flexibility of model-free learning and the robustness of model-based control. We evaluate performance both in simulation and in the real world.
  • Supplementary Material: zip
11 Replies