RoboHarness: Heterogeneous Policy Orchestration Boosts Long-Horizon Robotics to 95.2%
机器之心 · wechat · 2026-08-03
In embodied AI, control policies like VLA, RL, and TAMP each have distinct strengths, but a single model struggles with complex tasks requiring semantic understanding, geometric precision, and long-horizon reasoning simultaneously.
The paper introduces RoboHarness, a framework that shifts focus from waiting for a single general-purpose model to orchestrating heterogeneous policies. It solves two key challenges:
- Dynamic Policy Selection: Uses Understanding Skills to extract multi-dimensional metrics (e.g., semantic similarity, pose uncertainty) to dynamically characterize each policy's capability boundaries for sub-task routing.
- Bridging the "Handoff Gap": Addresses Out-of-Distribution (OOD) issues during policy switching. The Memory Bridge retrieves historical successful trajectories, explicitly reconstructs a familiar state distribution for the next policy, and generates bridging trajectories for stable handoffs.
Experiments show that in zero-shot evaluations on the long-horizon LIBERO-LoHo benchmark, RoboHarness achieves a complete success rate of 95.2%, significantly outperforming hierarchical decomposition methods using LLMs or world models (max 64.8%).
More from Embodied
- N_0-VTLA: First VTLA Foundation Model Pretrained on Tactile Data at Scale — NeoteAIEmbodied · 2026-08-03
- Foundation Robot Catches Baseballs with Open-Loop Control and Low-Friction Tendons — CyberRobooo · 2026-08-03
- Jensen Huang: Video Generation Models Have Solved the Physics Intuition for Robotics — r0ck3t23 · 2026-08-03
- OpenAI Hardware Roadmap Leaks: $200 Screenless Speaker and AI Phone Target 2027 — 创业邦 · 2026-08-03
- ACE-Data-0: New Embodied Data Engine on Hugging Face with 150 Hours of Multimodal Data — liuziwei7 · 2026-08-03
- Armory: Serve large robot policies to 10+ robots on a single GPU, boosting throughput by 18% — danfei_xu · 2026-08-03