EMHO: Self-Evolving Agent Harness Boosts Frozen Embodied Models via Experience Traces
Hyun Jung Lee · hf · 2026-10-14
Researchers from Korea University, U. Arkansas, and UW–Madison propose EMHO, a self-evolving framework that keeps the embodied model frozen and iteratively rewrites the surrounding agent harness—planning, context, and tool use—directly from execution traces under sparse environmental feedback.
Key points:
- EMHO goes beyond skills or recovery prompts, reshaping how the agent monitors progress, uses vision tools, grounds observations, and handles failures.
- EMHO-Merge uses episode-level gains and losses to decide when and how revised behaviors apply, letting one shared harness serve multiple subtasks.
- On EmbodiedBench navigation and manipulation tasks, EMHO consistently improves success rates for both Qwen 9B and 27B models, qualitatively changing how agents interpret and interact with their environment.
More from Embodied
- NAVER's WATCH: Cross-Video Human-Video Pretraining Lifts Robot Success Rate to 77.7% — naver-ai · 2026-10-14
- Robot Dog Tries to Climb Its Own Back: Conitzer's Latest Funny AI Fail — conitzer · 2026-10-14
- This Week in New Models: RoboJEPA, Long-WAM, VioLA and 4 More to Know — TheTuringPost · 2026-10-14
- XPENG's AUGUstus VLA model hits SOTA on NAVSIM and deploys on real vehicles at 12Hz — XPENG-AI · 2026-10-14
- Robotics startup Origin teases 'major announcement' with robot arm image — vaibhavbetter · 2026-10-14
- Paradigm appears to be getting into robotics with LeRobot arm kits — tensorqt · 2026-10-14