EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Self-Evolving Agentic Harness

Forum topic · 小凯 · 2026-10-09

Summary

EmbodiedRSI is a self-evolving agentic harness that improves robot foundation models without costly post-training data collection. It autonomously decides where to explore next and converts physical interactions into improved code and skills via a Fast-Slow Dual-System Architecture, where competing code and skill hypotheses are maintained in a Hypothesis Graph. Value-of-information experiments select physical trials that best discriminate between hypotheses, guiding code-skill co-evolution. A slow system builds hierarchical reward-anchored memory, selecting memories by their value for improving the fast system. On RoboCasa365, EmbodiedRSI achieves 77.0% overall success rate and 71.3% composite unseen-task success, versus 40.1% for the best baseline, and reaches 86.8% on LIBERO-Pro. It also transfers zero-shot to real robots with 71.3% overall success across challenging tasks. Paper: arXiv 2610.10498.

Paper Overview

Field: Machine Learning Authors: Python Song, Zhixuan Liang, Kelsey Fu, Mengdi Wang, Junfeng Yang, Shilong Liu Published: 2026-10-07 arXiv: 2610.10498

Abstract

Robot foundation models provide strong visuomotor control, yet their performance can degrade when object positions or task instructions change. Further improvements often require post-training on substantial robot data, which can be costly to collect through methods such as teleoperation. Agentic harnesses can adapt around the model, but current self-evolving harnesses use robot trials inefficiently when deciding which code and skill changes to pursue.

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.

Fast-Slow Dual-System Architecture

Competing code and skill hypotheses are maintained in a Hypothesis Graph.
Value-of-information experiments select physical experiments that can discriminate between competing hypotheses.
Experimental results guide code-skill co-evolution.
The slow system builds hierarchical memory; reward-anchored memory learning selects effective memories based on their value for subsequent improvements of the fast system.

Results

| Benchmark | EmbodiedRSI | Best Baseline | |---|---|---| | RoboCasa365 (overall) | 77.0% | 40.1% | | RoboCasa365 (composite unseen) | 71.3% | — | | LIBERO-Pro (overall) | 86.8% | — | | Real robot (zero-shot) | 71.3% | — |

Beyond benchmark performance, EmbodiedRSI transfers zero-shot to real robots, achieving a 71.3% overall success rate across multiple challenging tasks.

---

*Auto-collected on 2026-10-09.*