VLA 预训练中演示质量如何用：AXIS-Bench 基准与实证研究
What Matters in Quality-Aware VLA Pretraining: A Benchmark and Empirical Study
研究基于 5,000 条天然次优人类演示和 24 个目标任务，对比了 VLA 预训练中加权、过滤与条件化三种演示质量使用方式。结果显示，approach-aware 的分幕条件化在三个 VLA 上均优于普通行为克隆；条件化偏好分幕标签、过滤偏好分块标签，且存在模型相关的反转。团队提出 AXIS-Bench 支持可复现对比，真实实验里仿真预训练将成功率从 9.6% 提升到 31.7%。
Authors:Mingxuan Yan, Litian Gong, Dihong Huang, Zhixu Li, Tianyu Zhang, Wenqian Zhang, Hefeifei Jiang, Peihao Li, Jian Zhang, Mengfei Zhao, Yikai Tang, Hai Zhai, Zehao Wang, Ruijian Liang, Yanjia Huang, Lin Shao, Changliu Liu, Kaiyu Hang, Zhiwen Fan, Masayoshi Tomizuka, Jianfei Yang, Jiachen Li
Abstract:How should demonstration quality be used during vision-language-action (VLA) pretraining to improve performance after target-task fine-tuning? We compare weighting, filtering, and conditioning across reward definitions and label granularities using 5,000 naturally suboptimal human demonstrations and 24 target tasks. Within each VLA, we match demonstrations and training budgets. Additional comparisons examine pretraining data size and the roles of quality tags during fine-tuning and inference. Across three VLAs, we find that approach-aware episode conditioning improves mean success over plain behavioral cloning; conditioning and filtering favor episode and chunk labels, respectively, on average, with model-dependent reversals; and constant-tag fine-tuning improves mean success after conditioned pretraining with either tagged or untagged inference. We introduce AXIS-Bench to support these controlled comparisons. Built on crowdsourced simulation demonstrations, it provides recorded object and contact states for recomputing quality labels on shared trajectories. Replayable trajectories support workspace-aligned rendering and user-defined real-world target tasks. In our real-world experiment, simulation pretraining raises success from 9.6\% to 31.7\%. Project page: this https URL.
| Comments: | Project Website: this https URL |
| Subjects: | Robotics (cs.RO); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.18869 [cs.RO] |
| (or arXiv:2610.18869v1 [cs.RO] for this version) | |
| https://doi.org/10.48550/arXiv.2610.18869 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jiachen Li [view email]
[v1]
Tue, 13 Oct 2026 17:59:52 UTC (2,796 KB)
来源：arXiv：cs.AI · arxiv.org