Video World Models · 2026-11-11

WAM is capable in BEHAVIOR-1K: the 2nd-place solution to the 2026 BEHAVIOR Challenge

Yelin He, Jiaquan Ye, Xinming Wei, Wenjie Xiao, Gengming Zhang, Guoting Wei, Xianbiao Qi, Rong Xiao

arXiv:2611.13818PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors present the 2nd-place Team IF1001 solution to the 2026 BEHAVIOR Challenge (mobile manipulation over 100 household tasks). The method is built on Cosmos 3 as a world action model (WAM) and improved with richer visual history, task-progress estimation, task-group and skill-specific LoRA adapters, plus inference-time routing and execution strategies. The system achieves 0.615 mean Q-score and 41.2% full-task success on the official private test set.

Why it matters

The paper describes a WAM-based approach for long-horizon, mobile manipulation with non-Markovian ambiguity. It combines task-progress estimation and progress-conditioned routing to select task-group/skill policies, and reports official challenge metrics for BEHAVIOR-1K, including private-test performance under concurrency and limited hardware constraints.

Method

  • Use Cosmos 3 as the world action model backbone, then integrate richer visual history and explicit task-progress conditioning with stage ID inputs and task/skill text.
  • Train task-group LoRA adapters (15 groups of 6 tasks and 2 groups of 5) and skill-specific LoRA adapters (Pick/Place/Open/Close), with a progress estimator guiding routing and skill-to-stage enabling.
  • At inference, use task-progress estimation, action-candidate aggregation, rolling inpainting, task-specific execution windows/temporal retiming, and task-specific temporal and gripper control; avoid RSI/agentic harness to meet 1 FPS minimum.
Abstract (from arXiv)

This work presents the 2nd-place solution (Team IF1001) to the 2026 BEHAVIOR Challenge, which evaluates mobile manipulation across 100 household tasks. Our method is built on a powerful world action model (WAM), Cosmos 3. We improve Cosmos 3 by using richer visual history, task-progress conditioning, and task-group and skill-specific LoRA adapters. During inference, we apply a separate progress estimator to supply task-progress context, and use task- and progress-conditioned routing to select a task-group or skill-specific policy, complemented by action-candidate aggregation, rolling inpainting, and task-specific temporal and gripper control. To meet the minimum inference requirement of 1 FPS, we use neither recursive self-improvement (RSI) nor agentic harness methods and rely on no closed-source models; during official private-test evaluation, our system handled concurrent requests across 100 evaluation ports with limited hardware resources, sustaining an average end-to-end inference rate above 2 FPS per port. Our method achieves a mean Q-score of 0.613 and a full-task success rate of 40.9%, averaged over 3 self-evaluation rounds across all 100 tasks. On the official private test set, with 20 instances per task, our system achieves a mean Q-score of 0.615 and a full-task success rate of 41.2%.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First announced by arXiv: 2026-11-11 · Indexed by Pub-AI: 2026-11-11 UTC

preprint

WAM is capable in BEHAVIOR-1K: the 2nd-place solution to the 2026 BEHAVIOR Challenge

Yelin He, Jiaquan Ye, Xinming Wei, Wenjie Xiao, Gengming Zhang, Guoting Wei, Xianbiao Qi, Rong Xiao

First announced by arXiv: 2026-11-11 · Original record ↗

arxiv
2611.13818
Explicit source versions (1)
  • v1 · Date not supplied
  1. Preprint available

    Source date: 2026-11-11 · First detected: 2026-11-11 UTC · Historical baseline

    Source evidence (1)
    • Official evidence ↗ · RSS guid=oai:arXiv.org:2611.13818v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv

This work presents the 2nd-place solution (Team IF1001) to the 2026 BEHAVIOR Challenge, which evaluates mobile manipulation across 100 household tasks. Our method is built on a powerful world action model (WAM), Cosmos 3. We improve Cosmos 3 by using richer visual history, task-progress conditioning, and task-group and skill-specific LoRA adapters. During inference, we apply a separate progress estimator to supply task-progress context, and use task- and progress-conditioned routing to select a task-group or skill-specific policy, complemented by action-candidate aggregation, rolling inpainting, and task-specific temporal and gripper control. To meet the minimum inference requirement of 1 FPS, we use neither recursive self-improvement (RSI) nor agentic harness methods and rely on no closed-source models; during official private-test evaluation, our system handled concurrent requests across 100 evaluation ports with limited hardware resources, sustaining an average end-to-end inference rate above 2 FPS per port. Our method achieves a mean Q-score of 0.613 and a full-task success rate of 40.9%, averaged over 3 self-evaluation rounds across all 100 tasks. On the official private test set, with 20 instances per task, our system achieves a mean Q-score of 0.615 and a full-task success rate of 41.2%.