A humanoid on the monkey bars! 🐒 The robot from the @leggedrobotics Lab at ETH Zürich jumps up to the structure, swings across it, and drops down to a landing. 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. So the policy reads the raw scan from a head-mounted solid-state lidar, with no reconstruction step in between. An attention-based encoder picks out which of the sparse returns matter, and a GRU carries memory through the sequence, which counts when the geometry you're reaching for slides out of the sensor's view. Training is a phase-scheduled teacher-student setup. Separate privileged experts for jumping up, brachiating and jumping down, each with its own curriculum, and a scheduler handing control between them. The student is then distilled through DAgger, a critic warm-up, and PPO with the behaviour-cloning anchor decaying as reward takes over. The detail I enjoyed most sits in the sim-to-real section. Alongside LiDAR noise, they modeled battery voltage sag and actuator thermal limits. 🔗 Here's the paper page: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →