Noematrix open-sources RoboRSI, a multi-agent self-evolving harness for embodied AI

机器之心 · wechat · 2026-09-22

The article frames a scenario where frontier general models are taking over robot control (95% success on a manipulation task at 1/6 the token cost of a specialized model), and argues the real bottleneck is a system-level harness between understanding and safe, stable execution. Noematrix's RoboRSI addresses this:

Four-agent loop: Manager (scheduling, checkpoint recovery, skill versioning), Planner, Engineer (tool execution), and Reviewer (post-run diagnosis from logs/video/trajectories) form an execute→diagnose→revise→re-execute loop, with humans retaining control over goals and safety boundaries.
TSR skill tree: a four-layer hierarchy (task family → compound skill → atomic task → base skill) constrains coding agents so local fixes don't drift from global goals, turning history into reusable skills.
Code solidification: stable call chains are parameterized and frozen into code skills. On LIBERO, enabling it lifts pass rate from 21.5% to 29.0% while cutting median token use 29.4%, VLM calls 27.2%, and execution time 17.0%.
Results: after a day of parallel runs, zero-shot LIBERO pass rate rises from 32/120 to 95/120 and RoboTwin from 9/50 to 36/50; the full multi-agent config beats a single-Engineer baseline by 4x. On 840 LIBERO-Plus perturbations, adaptive revision lifts pass rate from 31.1% to 47.4%.

A comparable home-cleanup demo once required two engineers a month and 2-3k lines of code; RoboRSI closes the loop in one day, which the authors argue opens commercially viable personalized deployment scenarios.

Original post →

Noematrix open-sources RoboRSI, a multi-agent self-evolving harness for embodied AI

More from coding & agent