具身开发者社区

外观

论文arXiv 2610.1567517 天前

通过量化条件强化学习恢复视觉-语言-动作策略的低比特故障 ​

针对VLA量化导致任务完成下降，提出QCRL离线RL方法，并在真实世界试验与多种低比特设置上验证。

AI 整理
  • 硬件零部件
  • 具身模型
  • 端侧部署

原文链接：Recovering Low-Bit Failures in Vision-Language-Action Policies via Quantization-Conditioned Reinforcement Learning（arXiv 2610.15675 · 发布日期：2026-10-11）

Recovering Low-Bit Failures in Vision-Language-Action Policies via Quantization-Conditioned Reinforcement Learning

针对VLA量化导致任务完成下降，提出QCRL离线RL方法，并在真实世界试验与多种低比特设置上验证。

核心要点 ​

  • 量化可降低VLA部署成本，但PTQ/QAT在长程任务和视觉、语义偏移下仍落后FP。
  • QCRL结合进度优势估计与量化感知源条件，利用FP回放和部分完成信号。
  • 仅用同分布数据，π0.5的INT4 W4A16在三种视觉偏移下达到FP均分99.3%。
  • NVFP4 W4A4达到同一FP基线100.7%，NVFP4/FP8部署相对FP16提速1.73倍。

验证范围与局限：验证范围包括五个VLA模型、4050次真实世界试验，以及π0.5和GR00T N1.5的指定量化设置。

英文摘要 (Abstract) ​

Quantization lowers deployment costs of vision-language-action (VLA) policies but can compromise task completion. Across five VLA models and 4,050 real-world trials, we systematically reveal that quantization-aware training (QAT) generally improves on post-training quantization (PTQ) but leaves substantial gaps to full precision (FP) on long-horizon tasks and under visual and semantic shifts, while its imitation-based objective does not explicitly use rollout success or partial completion. To exploit these outcome signals, we introduce Quantization-Conditioned Reinforcement Learning (QCRL), an offline RL method coupling progress-based advantage estimation with quantization-aware source conditioning. FP rollouts supply reference experience beyond action-level distillation, progress-scaled terminal penalty distinguishes partial completion, and quantization-aware conditioning selects advantageous reference behavior within the quantized policy. We evaluate weight-only INT4 W4A16 on $π_{0.5}$ and GR00T N1.5 and joint weight-activation NVFP4 W4A4 on $π_{0.5}$. Using only in-distribution data, QCRL reaches 99.3% of FP's mean score across three visual-shift conditions for $π_{0.5}$ with INT4 W4A16, alongside 42.9% model-memory savings. With NVFP4 W4A4, it reaches 100.7% of the same FP baseline, alongside a $1.73\times$ speedup using NVFP4/FP8 deployment. Hardware gains are relative to FP16 on Jetson Thor.

作者：Seongmin Park, Jaewon Lee, Yujun Heo, Yoonsik Kim, Seheon Choi, Hyoung-Rock Kim, Jungwook Choi

来源

本页事实对应的公开页面，以原文为准。

  1. arXiv 2610.15675：Recovering Low-Bit Failures in Vision-Language-Action Policies via Quantization-Conditioned Reinforcement Learning访问 2026-10-21