AutoAgent0: VLM-powered agentic runtime lifts long-tail driving success by 28.3pp
yining_hong · x · 2026-10-16
Researchers introduce AutoAgent0, an agentic runtime for safe closed-loop autonomous driving. Instead of only optimizing end-to-end policies, it reframes runtime monitoring as an agentic harness that decides when to intervene and how to recover.
Key capabilities:
- Scene Context Refinement: VLM reasoning reconstructs critical scene context missed by nominal policies, verifying actions and triggering interventions on failure.
- Agentic Recovery Loop: VLM-guided stochastic optimization iteratively proposes, verifies, and refines explainable recovery policies.
Results: on Fail2Drive and NavSafe-∞, the framework improves closed-loop robustness across multiple frozen driving experts — 81.0% long-tail success rate (+28.3pp over SparseDriveV2), 90.3% average success on BridgeDrive, and gains on all 3 Fail2Drive experts, at some cost to comfort and efficiency. Paper, code, and demos are available.
More from Embodied
- DynaRobotics CEO Lindon Gao on running deployment and research feedback loops in parallel — micoolcho · 2026-10-16
- Baidu to give 1,000+ Hong Kong residents free robotaxi rides in city-first public trial — Baidu_Inc · 2026-10-16
- Auki demos spatial harness: robots and devices share spatial context via AR — broodsugar · 2026-10-16
- Tesla's 2-DOF car becomes 6-DOF flying demo — easier control problem than humanoids — micoolcho · 2026-10-16
- XREAL Aura, Android XR glasses powered by Gemini, named a TIME Best Invention of 2026 — DuRuofei · 2026-10-16
- CarryOn: dual-loop harness adds personalization and safe recovery to frozen VLA robots — Zixuan Wang · 2026-10-16