ETH Zurich humanoid swings across monkey bars on raw lidar
The policy skips the terrain map entirely and feeds sparse lidar returns straight into an attention encoder.
A humanoid from the Legged Robotics Lab at ETH Zurich jumps up to a set of monkey bars, swings across them, and drops down to a landing. Lukas Ziegler wrote up the paper and flagged the part that matters most for anyone building legged controllers.
Most legged robots build a terrain map first, and a heightfield is the wrong representation for a horizontal bar with air above and below it. Thin overhanging geometry is exactly what a map throws away.
The detail I enjoyed most sits in the sim-to-real section. Alongside LiDAR noise, they modeled battery voltage sag and actuator thermal limits.
No map
Most legged robots build a terrain map before they move. Ziegler's point is that a heightfield, the usual map format, is the wrong shape for a horizontal bar with air above and below it. Thin overhanging geometry is exactly what a map deletes.
So this policy skips the reconstruction step. It reads the raw scan from a head-mounted solid-state lidar directly. An attention-based encoder decides which of the sparse returns actually matter, and a GRU (a small recurrent memory) carries state through the sequence, which counts when the bar you are reaching for slides out of the sensor's field of view mid-swing.
How it was trained
The setup is phase-scheduled teacher-student. Separate privileged expert policies handle jumping up, brachiating (swinging hand over hand along the bars) and jumping down, each with its own curriculum, with a scheduler passing control between them. Those experts are then distilled into one student through DAgger (the teacher corrects the student's mistakes as it runs), a critic warm-up, and PPO (a reinforcement learning algorithm), with the behaviour-cloning anchor decaying as reward takes over.
The detail Ziegler singled out sits in the sim-to-real section (trained in simulation, then run on the real robot). Alongside lidar noise, the team modeled battery voltage sag and actuator thermal limits. Those are the unglamorous terms that decide whether a policy that looks clean in simulation still works when a real robot's battery droops under a hard swing and its motors heat up.
What is not in the post
Ziegler's writeup does not give a success rate, a number of attempts, or how many bars the robot crosses in a run. There is no claim about how often it falls. Nobody outside the lab has reported reproducing it.
The transferable idea here is the sensing choice rather than the stunt. Monkey bars are a neat showcase, but the reason other teams will read the paper is that it argues raw lidar into the policy beats a reconstructed map when the thing you need to grab is thinner than the map's resolution. Ledges, railings, scaffolding and pipes have the same problem.