Automatic discovery of diverse whole-body contact strategies in higher-DoF robots

Determine whether the AME-2 reinforcement learning framework with attention-based neural map encoding, trained without additional references or priors, can automatically discover multiple distinct whole-body contact strategies in higher-degree-of-freedom robots such as humanoids, enabling the policy to select between behaviors like single-leg stepping, two-leg jumping, and combined arm–leg contacts to handle the same terrain type at different scales.

Background

The paper demonstrates emergent whole-body contact skills but notes that learned motions tend to follow similar contact patterns for a given terrain type. For higher-DoF systems (e.g., humanoids), the same terrain class at different scales may require qualitatively different contact strategies, such as stepping, jumping, or using arms.

The authors explicitly state uncertainty about whether their AME-2 method, without leveraging additional priors or reference motions, can autonomously discover such a diverse set of strategies to cover a broader range of terrain scales.