arXiv cs.AI· Hongzhan Yu, Chenghao Li, Ruipeng Zhang, Henrik Christensen, Sicun Gao·· 10 天前AI 评分22
面向潜在动力学模型的敏感度塑形：提升 OOD 检测与闭环规划安全性
Sensitivity Shaping for Latent Modeling
AI 导读
研究提出 support-conditioned control-sensitivity regularization，通过在被充分支持训练区域增强局部响应性，保留控制变化引起的潜在状态变化，解决学习动力学对控制不敏感、将无支持控制映射为类似已演示转移的潜在预测、从而抑制 OOD 信号的问题。
正文 · 原文
Abstract:Generative dynamics models enable planning in challenging systems, but safe deployment requires detecting policy-induced out-of-distribution (OOD) transitions. Existing methods typically treat learned dynamics as fixed and rely on post hoc support surrogates for OOD detection. This overlooks a critical failure mode: learned dynamics that are insensitive to control changes can map unsupported controls to latent predictions resembling demonstrated transitions, suppressing OOD signals despite large prediction errors. We introduce support-conditioned control-sensitivity regularization to preserve control-induced variation by promoting local responsiveness in well-supported training regions. Experiments in vision-based obstacle avoidance, manipulation, and real-robot navigation demonstrate improved OOD detection and safer closed-loop planning.
| Comments: | Conference on Robot Learning (CoRL) 2026 |
| Subjects: | Robotics (cs.RO); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2606.14585 [cs.RO] |
| (or arXiv:2606.14585v2 [cs.RO] for this version) | |
| https://doi.org/10.48550/arXiv.2606.14585 arXiv-issued DOI via DataCite |
Submission history
From: Hongzhan Yu [view email]
[v1]
Fri, 12 Jun 2026 16:01:50 UTC (23,027 KB)
[v2]
Tue, 6 Oct 2026 17:21:45 UTC (23,035 KB)
来源：arXiv cs.AI · arxiv.org