Automatic Discovery of Diverse Whole-Body Contact Strategies

Diversity-Augmented RL for humanoid locomotion: discovering multiple contact strategies without reference motions or priors.

4.7
Avg Strategies Discovered (DARL)
87%
Terrain Coverage
82%
Success Rate
3.9x
More Strategies vs Baseline

Discovered Contact Strategies

Single-Leg Step Two-Leg Jump Arm-Leg Combined Crawl Shuffle Vault

Strategies per Terrain Type

Method Comparison

Performance Comparison

MethodCoverageAvg SpeedSuccess RateStrategies
Baseline RL62%1.15 m/s58%1.2
QD Only78%0.92 m/s71%3.8
Diversity Reward73%1.08 m/s67%3.2
DARL (Full)87%1.05 m/s82%4.7
DARL achieves the best terrain coverage (87%) and success rate (82%) with only a modest 9% speed reduction vs baseline. The combination of QD archive and diversity rewards is essential: each component alone provides partial benefits but the synergy achieves the best results.
Diverse Whole-Body Contact Strategies — Computational Study