HALO: Lyapunov-Certified MARL Brings Stable Human-Robot Collaboration to Real Humanoids, ICML 2026 Oral
rohanpaul_ai · x · 2026-10-30
Researchers from UT Arlington and CMU propose HALO (Heterogeneous-Agent Lyapunov policy Optimization), an ICML 2026 Oral addressing the 'rationality gap' in human-robot collaboration: the mismatch between decentralized best-response dynamics and centralized cooperative ascent makes independent policy-gradient updates oscillate or diverge in general-sum differentiable games.
HALO enforces a per-step Lyapunov decrease condition on a parameter-space disagreement metric, correcting decentralized gradients via optimal quadratic projections to guarantee monotonic contraction of the rationality gap. Unlike Lyapunov-based safe RL targeting state/trajectory constraints, it uses Lyapunov certification to stabilize decentralized policy learning itself.
Simulations and real-world humanoid experiments across three collaborative tasks (OSP, SCT, SLH) show improved generalization and robustness in corner cases, with sim-to-real deployment. Paper, code, and poster are publicly available.
More from Embodied
- Noninvasive BCI lets unresponsive ICU patients communicate via brainwaves — ZeroStateReflex · 2026-10-30
- Musk says Neuralink's Blindsight is engineering-ready, awaiting regulatory approval — elonmusk · 2026-10-30
- Cute wheeled desktop robot serves drinks at LongJourney VC demo night — aniketmaurya · 2026-10-30
- Tiny M5Stack robot fangirls over Teknium in English and Japanese via LFM3 Audio — alexcovo_eth · 2026-10-30
- Mystery agent method tops RoboDojo at 73.33, first past 70 points and 12 points ahead of runner-up — jiqizhixin · 2026-10-30
- Artificial skin lets robot fingers grasp pencil lead and dodge needles, reading 10,000+ sensors 100x/sec — CurieuxExplorer · 2026-10-30