跳转至

FACT：世界-动作模型的故障感知因果训练

文章背景与核心概要

世界-动作模型（WAMs）在具身智能和机器人领域展现出巨大的潜力，但它们通常面临一个核心痛点：模型主要基于成功的演示进行训练，导致其对“错误动作”引发的后果缺乏认知。为了解决这一局限，研究人员提出了 FACT（Failure-Aware Causal Training，故障感知因果训练）方法。这是一种因果世界-动作模型，能够将未来的视频和任务进展预测条件化于实际执行的动作之上。

通过将失败轨迹（failure rollouts）引入训练流程，FACT 能够监督错误动作的后果，将原本会被丢弃的失败数据转化为有效的未来预测目标，从而显着减少受成功偏见（success-biased）影响的未来幻觉。在模拟环境以及真实世界的双臂操作任务上的广泛测试表明，FACT 的性能显著优于传统基线，并且随着更多失败数据的融入，其性能会进一步提升。

摘要 (Summary)

World-action models (WAMs) often struggle because they are primarily trained on successful demonstrations, leaving them blind to the consequences of bad actions. To solve this, researchers introduce FACT (Failure-Aware Causal Training), a causal world-action model that conditions future video and task progress predictions on executed actions. By incorporating failure rollouts into training, FACT supervises action consequences, turns bad actions into valid future targets, and significantly reduces success-biased future hallucinations. Extensive testing on simulation and real-world bimanual manipulation tasks shows that FACT outperforms traditional baselines and improves as more failure data is integrated.

世界-动作模型（WAMs）往往存在性能瓶颈，因为它们主要基于成功的演示进行训练，从而对坏动作（bad actions）导致的后果盲目。为了解决这一问题，研究人员引入了 FACT（故障感知因果训练，Failure-Aware Causal Training），这是一种因果世界-动作模型，它将未来的视频和任务进展预测条件化于所执行的动作上。通过将失败回滚（failure rollouts）纳入训练，FACT 监督了动作的后果，将坏动作转化为有效的未来目标而不是直接丢弃。在模拟和真实世界的双臂操作任务上进行的广泛测试表明，FACT 优于传统的基线，并且随着更多失败数据的整合而不断改进。

论文元数据 (Paper Metadata)

  • arXiv Identifier: arXiv:2608.10232 [cs.RO]
  • Primary Subject: Robotics (cs.RO)
  • Secondary Subjects: Artificial Intelligence (cs.AI), Machine Learning (cs.LG)
  • Submission Date: August 10, 2026
  • Authors: Quanquan Peng, Yutong Liang, Rui Yan, Nicklas Hansen, Xiaolong Wang
  • Project Website: fact-wam.github.io
  • arXiv Identifier: arXiv:2608.10232 [cs.RO]
  • Primary Subject: Robotics (cs.RO)
  • Secondary Subjects: Artificial Intelligence (cs.AI), Machine Learning (cs.LG)
  • Submission Date: August 10, 2026
  • Authors: Quanquan Peng, Yutong Liang, Rui Yan, Nicklas Hansen, Xiaolong Wang
  • Project Website: fact-wam.github.io

摘要正文 (Abstract)

Recent world-action models (WAMs) show that co-training policies with future prediction can provide physical priors for action generation. Building on the future-prediction ability of video models, many WAMs generate future videos and recover actions with inverse-dynamics models, or use these predicted videos as goal conditions for action generation. In both cases, the world model is trained mostly on successful demonstrations and has little reason to predict the consequences of bad actions.

最近的世界-动作模型（WAMs）表明，将策略与未来预测进行联合训练可以为动作生成提供物理先验。基于视频模型的未来预测能力，许多 WAM 生成未来视频并通过逆动力学模型恢复动作，或者将这些预测的视频用作动作生成的具象目标条件。在这两种情况下，世界模型大多基于成功演示进行训练，因此几乎没有动机去预测错误动作的后果。

We introduce FACT, a causal World-Action Model that predicts future video and task progress conditioned on the executed action. This action-conditioned interface allows failure rollouts to supervise action consequences, turning bad actions into valid future targets rather than being discarded. Failure-aware training makes the progress predictor aware of both successful and failed action outcomes, which can optionally be used to score sampled action candidates at inference. Extensive experiments on simulation and real-world bimanual manipulation tasks show that FACT outperforms many existing baselines, improves as failure data are incorporated into training, and reduces success-biased future hallucination under bad actions.

我们引入了 FACT，这是一种因果世界-动作模型，它根据执行的动作为条件预测未来的视频和任务进展。这种动作条件化接口允许失败回滚来监督动作后果，将不好的动作转化为有效的未来目标，而不是被丢弃。故障感知训练使进展预测器能够意识到成功和失败的动作结果，这在推理时可以可选地用于对采样的动作候选进行评分。在模拟和真实世界的双臂操作任务上的大量实验表明，FACT 优于许多现有的基线，随着失败数据被纳入训练而不断改进，并减少了在坏动作下受成功偏见影响的未来幻觉。