Agile and Generalized Legged Locomotion

via Attention-Based Neural Map Encoding

IEEE TRO 2026

Authors: Chong Zhang, Victor Klemm, Fan Yang, Marco Hutter

ETH Zurich

Paper Link: https://arxiv.org/abs/2601.08485

aka: AME-2. Tuning Guide: Link

Youtube: https://www.youtube.com/watch?v=FNA1crvtBLs

Code Release: https://github.com/leggedrobotics/ame2_minimal

Highlights

journal_video_v3.mp4

more_results_final.mp4

More zero-shot results in simulation & Baseline Comparison

simzs.mp4

simcompare.mp4

Works for Humanoids (e.g., G1, H1)

all_terrains_4x3.mp4

G1↑: extra RBF-based shaping for upper body pose and ankle yaw

H1→: same rewards as ANYmal-D and TRON1

H1results.mp4

After minor addition of regularization, it works nicely on real G1.

Method

Policy Training

Policy Architecture

Neural Mapping

Prior Work

He, Junzhe, Chong Zhang, Fabian Jenelten, Ruben Grandia, Moritz Bächer, and Marco Hutter. "Attention-based map encoding for learning generalized legged locomotion." Science Robotics 10, no. 105 (2025): eadv3604.

Page updated

Report abuse