ETH Zurich: foothold prediction cuts humanoid locomotion training to ~1/3 iterations
breadli428 · x · 2026-10-28
A paper from ETH Zurich's Robotic Systems Lab (Marco Hutter's group), "Watch Your Step: Anticipated Perception for Humanoid Locomotion," replaces dense elevation-map inputs with predicted footholds.
- A lightweight foothold prediction model (FPM) forecasts future footholds (two steps ahead) from proprioception and velocity command alone
- The policy only sees small elevation-map patches around predicted footholds, removing redundant inputs
- FPM trains alongside the locomotion policy, periodically retrained on fresh rollouts
Results: reaching a given terrain difficulty takes roughly 1/3 the training iterations vs. the full map; it beats full-map policies on sparse-foothold terrain (including narrow walled corridors) and unseen terrain; matches an attention-based encoder trained at several times the cost. Zero-shot transfer to a Unitree G1 crossing obstacles, stairs, gaps, and step-downs. Paper and code are public.
More from Embodied
- Sony Launches ELF-MN1, a 6-inch Glasses-free Spatial Display for Desktop Avatars — TinfoilTricorn · 2026-10-29
- Foundry Robotics raises $100M Series A to build 130,000 sq ft factory in Emeryville, California — ishabytes · 2026-10-29
- NVIDIA's Metatron lets coding agents drive humanoid whole-body motion via a goal interface — carlosdponx · 2026-10-29
- StackChan Desk Robot Runs Grok + Hermes Agent, Lights Sync With Mood — alexcovo_eth · 2026-10-29
- Purdue's GPS-free drone tracks runners through forests; Rachel Thomas asks if it should exist — mmitchell_ai · 2026-10-29
- Insta360 X6 Spatial Capture Turns a 360° Walk Into an Explorable Gaussian Splat Scene — Aiden_Tech_Ai · 2026-10-28