AI TrackerFrontier desk · AEST
Back to news
normalaitech

EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution

arXiv cs.AI · 8 Oct 2026, 03:48 AEST

EmbodiedRSI is a self-evolving agentic harness that enhances robot learning through hypothesis-guided co-evolution. It achieves significant performance improvements in various robotic tasks compared to existing methods.

Key facts

Read aloud

EmbodiedRSI is a self-evolving agentic harness that enhances robot learning through hypothesis-guided co-evolution. It achieves significant performance improvements in various robotic tasks compared to existing methods. Robot foundation models provide strong visuomotor control, yet their performance can degrade when object positions or task instructions change. We introduce EmbodiedRSI, a self-evolving agentic harness that autonomously decides where to explore next and turns the resulting physical interaction into improved code and skills. On RoboCasa365, EmbodiedRSI reaches seventy seven point zero percent overall success and seventy one point three percent on Composite-Unseen, compared with forty point one percent for the best baseline. EmbodiedRSI also reaches eighty six point eight percent overall success on LIBERO-Pro. Beyond benchmark performance, EmbodiedRSI transfers zero-shot to real-world robot, achieving seventy one point three percent overall success across multiple challenging tasks.

Sources in this cluster