RoboFoundry freezes the model, evolves the system: GPT-5.5 embodied success jumps to 72.7%

机器之心 · wechat · 2026-10-12

Researchers from NTU, Beihang, NUS, SJTU and others propose RoboFoundry, the first framework formalizing 'Self-Evolving System-as-Policy' for embodied agents: the base model stays frozen while the surrounding context system, skill system, and execution interface evolve. The system lives as files the model edits like a codebase; an execute-reflect-repair-promote loop attributes failures to six capabilities, and only repairs validated on the current task plus replay-checked on other tasks get promoted into the shared general system, with an auditable 'Evolution Ledger.' Results: on EmbodiedBench all five base models improve—GPT-5.5 jumps from 56.9% to 72.7% (beating un-evolved GPT-6 Astra at 71.9%) and the spread between bases shrinks from 19.6 to 11.5 points; RoboMemArena TSR beats all baselines by ≥39.0%; LIBERO-PRO relative gains of 243.8–679.7% over CaP-Agent0. Real-robot tests (GPT-5.5 + π0.5 backend): online evolution lifts nesting-doll success to 70% vs 30%; towel-folding transfers zero-shot to unseen towels (>80% vs <20%); plus semantic navigation on Unitree G1 and a six-step chemistry experiment with long-horizon state tracking. The takeaway: embodied progress comes not just from stronger models but from continuously strengthening the system around them.

Original post →

More from Embodied

Embodied channel →