RAI Institute's RL bicycle robot nails table jumps, flips and wheelies
KyleMorgenstein · x · 2026-11-15
RAI Institute researcher Shamel Fahmi (featured among "AI Native" companies at NVIDIA's GTC keynote) shared the paper "Bicycle Acrobatics with Reinforcement Learning":
- Bicycle robots are fast and efficient but underactuated, non-holonomic dynamics make agile maneuvers hard
- Multiple RL formulations (waypoint following, pose reaching, twist tracking, motion imitation) teach the Ultra Mobility Vehicle forward/lateral jumps, steerable jumps, front flips, kip-ups, wheelies and more
- An orchestrator switches between policies via state-dependent triggers for long-horizon stunts
- Hardware-validated: repeated table traversals up to 1 m, 15+ consecutive autonomous jumps, 20+ consecutive stunt repertoires
Result: RL endows wheeled robots with legged-platform agility while keeping the speed and efficiency of wheels.
More from Embodied
- SpaceX Appears to Catch an Orbital Upper Stage Mid-Air — VraserX · 2026-11-15
- Ex-Apple exec brings a full-size humanoid robot into the office break room — cixliv · 2026-11-15
- Turing winner Patterson: general AI models will do most jobs by end of next year — davidpattersonx · 2026-11-15
- ROS 2 Record-and-Replay for the SO-101 Arm Using MCAP and Foxglove — kamathsblog · 2026-11-15
- He reverse-engineered his smart glasses into a live AI object-recognition rig with Opus and YOLO — IgorCarron · 2026-11-15
- SJTU's Value-as-Memory Gives Robot Policies Task-Progress Memory, +29.4pp on Real Tasks — jiqizhixin · 2026-11-15