提出跨范式微调方法，让视觉语言动作模型在连续动作上也能高效优化。
CrossVLA: Cross-Paradigm Post-Training and Inference Optimization for Vision-Language-Action Models
- 设计新概率估计器，避免积分即可对连续动作模型做偏好优化
- DoRA比LoRA提升10.4个百分点，多任务测试中表现稳定且无种子差异
- 发现推理时去噪循环占78.6%耗时，缓存策略加速有限且可能降低成功率
视觉-语言-动作（VLA）模型已趋于统一架构：离散标记自回归（如OpenVLA）与连续动作流匹配（如pi-0.5）。然而，直接偏好优化（DPO）作为语言模型的主流后训练方法，几乎仅研究于自回归VLA。本文提出CrossVLA，开展跨范式VLA后训练的实证研究。三项贡献：(i) 提出替代流匹配的概率估计器，使DPO无需概率流常微分方程求解即可作用于连续动作主干；(ii) 头对头比较LoRA与DoRA作为参数高效层在VLA DPO中的表现，发现DoRA相较OpenVLA SFT在LIBERO 4套件上平均提升+10.4个百分点（600次试验，3个种子），各子任务分别提升+20.0（物体）、+11.0（长序列）、+8.0（目标）、+2.7（空间），且物体任务在三种子下均达38/50，无种子方差；(iii) 推理阶段分析显示，去噪循环占据78.6%的采样耗时，前缀键值缓存类VLA-Cache最多仅能带来21%加速，块级与标记级缓存策略均导致成功率下降至0-80%。此外，我们在6000张LIBERO图像上预训练了多视角+时间投影头，实现99.5% k-NN召回@1（较随机高36倍），可作为下游初始化。所有代码、检查点、训练日志及复现脚本开源于https://github.com/lz-googlefycy/vla-lab。
原文摘要 · Abstract (English)
Vision-Language-Action (VLA) models have rapidly converged on a small set of architectural patterns: discrete-token autoregression (e.g. OpenVLA) and continuous-action flow-matching (e.g. pi-0.5). Yet preference alignment via Direct Preference Optimisation (DPO) -- the de-facto post-training step in language models -- has been studied almost exclusively on autoregressive VLAs. We present CrossVLA, an empirical study of cross-paradigm VLA post-training. Three contributions: (i) a surrogate flow-matching log-probability estimator that lets DPO operate on continuous-action backbones without probability-flow ODE integration; (ii) a head-to-head comparison of LoRA and DoRA as the parameter-efficient layer for VLA DPO, finding DoRA improves over OpenVLA SFT by a mean +10.4 pp across LIBERO 4-suite (600 trials, 3 seeds) -- per-suite +20.0 Object, +11.0 Long-horizon, +8.0 Goal, +2.7 Spatial -- with zero seed variance on Object (38/50 on each of 3 seeds); (iii) an inference-time anatomy showing the denoise loop dominates 78.6% of sample_actions latency and prefix-K/V caching a la VLA-Cache caps at a 21% acceleration ceiling -- both chunk-level and token-level cache strategies degrade success rate to 0-80% in our benchmarks. We further pretrain a multi-view + temporal projection head on 6000 LIBERO frames, achieving 99.5% k-NN recall@1 for same-task retrieval (36x over random), available as a downstream initialisation. All code, ckpts, training logs, and reproduction scripts are open at https://github.com/lz-googlefycy/vla-lab.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。