Today for AI
HOT RADAR
agentHEAT 10.1°

AI CLUSTERED EVENT · 10/12/2026

RoboFoundry Introduces Self-Evolving System-as-Policy Framework for Embodied Agents via Frozen Base Models

1 reports archived1 independent sourcesupdated 10/12/2026, 12:30:00
Synthesis & Latest Updates
1 Sources Cross-Validated

Researchers from NTU and other institutions introduced RoboFoundry, formalizing the entire agent system as a 'Self-Evolving System-as-Policy'. By keeping base models frozen and autonomously evolving context, skills, and execution interfaces, it boosted GPT-5.5's success rate on EmbodiedBench from 56.9% to 72.7%, significantly outperforming baselines in long-term memory and perturbation tests.

LATEST/Researchers from NTU and other institutions introduced RoboFoundry, formalizing the entire agent system as a 'Self-Evolving System-as-Policy'. By keeping base models frozen and autonomously evolving context, skills, and execution interfaces, it boosted GPT-5.5's success rate on EmbodiedBench from 56.9% to 72.7%, significantly outperforming baselines in long-term memory and perturbation tests.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. 机器之心 (微信公众号)T2·86 pts
    • New Paradigm: Defines 'System-as-Policy', treating context, skills, and execution interfaces as independently evolvable components rather than static configurations.
    • Zero-Shot Transfer: Demonstrates cross-robot and cross-task zero-shot transfer capabilities along with online evolution after deployment on real robots.
    • Significant Performance Boost: Improved GPT-5.5's EmbodiedBench success rate by nearly 16 percentage points without parameter changes; outperformed CaP-Agent0 by 243.8–679.7% in LIBERO-PRO perturbation tests.