Huawei Noah wins BEHAVIOR-1K Challenge with 48% task success rate, 4x last year
机器之心 · wechat · 2026-11-14
Huawei Noah's TheNorthStar team won the BEHAVIOR-1K Challenge 2026 among 57 global teams, achieving 0.683 Q-score and 48.2% full task success on 100 hidden household tasks — 4x last year's best (12%). The benchmark, built on Fei-Fei Li's BEHAVIOR-1K and tied to CoRL 2026, tests long-horizon reasoning, cross-room navigation, and bimanual manipulation.
Technical report: Scaling by Structure
- Their HiPhys-RSI framework uses a frozen coding agent plus VLM to accumulate interaction experience into executable programs: agents read logs, analyze videos, localize failures, and iteratively refine pipelines from high-level orchestration down to tool implementations.
- Validated programs become reusable components; at inference a "capability card" selects pipelines or VLA policies.
- The hybrid system outscored three VLA training baselines on nearly all 100 tasks across all eight skill categories.
- Improvement follows three phases: early gains from combining existing capabilities, a plateau, then late gains from discovering novel manipulation strategies.
- Pipeline-generated trajectories can also feed VLA post-training, enabling co-evolution between structured knowledge and learned policies.
Grasping, target search, and timeouts remain dominant failure modes; the report also discusses evaluation loopholes exposed by strategy exploration.
More from Embodied
- Ropedia open-sources Xperience 2: 1,589 3D scene scans and 131 hours of egocentric demos for physical AI — liuziwei7 · 2026-11-14
- RobotAgentBench: hybrid Codex-InternVLA-A1.5 system averages 48.5 on household agent tasks — jiqizhixin · 2026-11-14
- ASU IRIS Lab recruiting fully funded PhD students for Fall 2027 in contact-rich dexterity — YuXiang_IRVL · 2026-11-14
- OpenAI disbands robotics team 5 months after relaunch as Meta swoops in hiring — svlevine · 2026-11-14
- CoRL 2026: a robot shot my Polaroid, a Cybercab drove me around — what evidence would convince you? — bousmalis · 2026-11-14
- One engineer builds complete AI smart glasses from PCB to finished product — IgorCarron · 2026-11-14