ETH Zurich's Humanoid Robot Swings Across Monkey Bars in 14 of 15 Tries
A humanoid robot jumps onto monkey bars, swings across at half a meter per second, and drops safely, guided by raw lidar and a reinforcement-learned policy.
PRO- ETH team trains a humanoid to jump onto monkey bars, swing across, and land, succeeding in 14 of 15 trials.
- Policy reads raw solid-state lidar through an attention encoder with recurrent memory, no elevation map required.
- Three privileged experts for jump-up, brachiation, jump-down are distilled into one student under a phase schedule.
- Sim-to-real models lidar noise, battery-voltage sag, and motor thermal limits; hardware uses passive hook end-effectors.
- Reaches brachiation speeds up to 0.5 m/s on real hardware across three bar configurations.
- Same perception backbone generalizes to ducking under 2 cm overhead obstacles, suggesting a reusable sparse-3D perception module.
Getting a humanoid robot to walk over rough ground is one thing. Leaping onto monkey bars, swinging hand over hand, and dismounting cleanly is an entirely different class of problem. A new paper from ETH Zurich tackles exactly that, with the resulting policy completing the full sequence in 14 of 15 hardware trials.
The work, titled Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids, uses monkey-bar brachiation as a stress test for three problems that usually get studied in isolation: perceiving thin geometry with lidar, discovering precise contact sequences through exploration, and executing whole-body motion within the torque and thermal limits of real hardware.
Why monkey bars are a nasty benchmark
Most perceptive locomotion research assumes the robot stands on something. Ground can be represented as an elevation map, a 2.5D grid where every cell has a height. Monkey bars break that assumption. The support structure is overhead, sparse, and made of thin cylinders that solid-state lidar barely registers. Adjacent work like APEX (high-platform climbing) and perceptive humanoid parkour has pushed on parts of this problem, but none combined a mid-air jump, a swinging traverse, and a controlled landing in a single policy.
Raw lidar straight into the policy
The robot carries a head-mounted solid-state lidar. Because the bars are thin, each scan produces only a handful of relevant returns, and those returns jitter frame to frame. Rather than pre-processing into a voxel grid or fitting geometry, the policy consumes raw lidar through an attention-based encoder with recurrent memory. The attention layer picks out sparse task-relevant points from thousands of returns, while the recurrent layer integrates them over time so the robot accumulates a stable estimate of each bar's position even when individual scans are noisy. The encoder design draws from prior attention-based work on elevation-map encoding for legged locomotion, adapted here to operate on raw point returns instead of a pre-built map.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
