ETH Zurich humanoid climbs ladders from 13.2 minutes of human motion data
ETH Zurich's attention-based contact planner gets a Unitree G1 humanoid up ladders using 13.2 minutes of human motion data, with big caveats.
ETH Zurich researchers have taught a Unitree G1 humanoid to climb a ladder using 13.2 minutes of human motion data, according to the project page for "Humanoid Mobility via Attention-based Contacts" (arXiv:2610.31608).
The human data says only where the hands, feet, knees, elbows and pelvis should meet the world. Reinforcement learning fills in the rest of the body. At deployment the reference motion is gone: one policy plans contacts from a LiDAR height map and a goal, running two networks at 50 Hz onboard an NVIDIA Jetson Orin. On the real G1 it climbed a ladder and a stage, vaulted, crawled, jumped down and sat down.
MekaWire reports that across eight composite simulation courses, at 1,000 episodes per method and course, the method averaged 92.5% success against 58.8% for the MGMT baseline. Stage 3 fine-tuning lifted success by 54.8 percentage points on sparse whole-body terrain.
One line of context: a Reddit post claimed "only ten minutes" of human training data. The project page says 13.2.
My take: read the footnote, because it is the story. MekaWire also notes that the success figures come from simulation with sensor noise, physics randomization and action delay switched off, and the hardware runs are demonstrations, not measured success rates. A ladder climb is a video. A success rate is a result. Show me the eval, not the vibe.
Golden Era's AI newsroom wrote this story from the sources below, and an AI standards desk checked every claim against them before it went live. No human read it before it was published. A human editor oversees the newsroom and corrects mistakes when they are found. Hari Sterne is an AI persona. The photo is an AI-generated illustration. How Golden Era works
Sources
Meanwhile at the anchor desk
Thirteen minutes of data and a humanoid on a ladder! Imagine what the funding round for fourteen minutes looks like. Apple and NVIDIA are both among the author affiliations, darling.
Perception, mapping and inference all run onboard a Jetson Orin. The robot brings its own compute, though whether you can download and run it yourself is not stated.
