LaST-R1: Reinforcing Action via Adaptive Physical Latent Reasoning for Vision-Language-Action Models

Forum topic · 小凯 · 2026-05-03

Summary

LaST-R1 is a unified Vision-Language-Action (VLA) framework for robotic manipulation that integrates latent Chain-of-Thought (CoT) reasoning over physical dynamics before action execution, paired with a dedicated reinforcement learning post-training paradigm. Existing VLA approaches either rely on explicit linguistic reasoning, which suffers from latency and discretization issues, or use continuous latent reasoning but remain confined to static imitation learning, limiting adaptability. LaST-R1 introduces two key contributions: (1) Latent-to-Action Policy Optimization (LAPO), a novel RL algorithm that jointly optimizes the latent reasoning process and action generation, improving physical world modeling and robustness in interactive environments; and (2) an adaptive latent CoT mechanism that dynamically adjusts reasoning depth based on environmental complexity. Experiments show LaST-R1 achieves a near-perfect 99.8% average success rate on the LIBERO benchmark with only one supervised warm-up stage, significantly outperforming prior state-of-the-art methods in both convergence speed and performance. In real-world deployment, LAPO post-training improves over the initial policy by up to 44% across four complex tasks, including single-arm and dual-arm setups. Paper: arXiv 2604.28192.

Paper Overview

Field: Computer Vision (CV) Authors: Hao Chen, Jiaming Liu, Zhonghao Yan, Nuowei Han, Renrui Zhang, Chenyang Gu, Jialin Gao, Ziyu Guo, Siyuan Qian, Yinxi Wang, Peng Jia, Chi-Wing Fu, Shanghang Zhang, Pheng-Ann Heng Published: 2026-04-30 arXiv: 2604.28192

Abstract

Vision-Language-Action (VLA) models have increasingly incorporated reasoning mechanisms for complex robotic manipulation. However, existing approaches share a critical limitation: whether employing explicit linguistic reasoning that suffers from latency and discretization, or utilizing more expressive continuous latent reasoning, they are predominantly confined to static imitation learning that limits adaptability and generalization. While online reinforcement learning (RL) has been introduced to VLAs to enable trial-and-error exploration, current methods exclusively optimize the vanilla action space, bypassing the underlying physical reasoning process.

Key Contributions

This paper presents LaST-R1, a unified VLA framework that integrates latent Chain-of-Thought (CoT) reasoning over physical dynamics before action execution, accompanied by a specialized RL post-training paradigm:

1. Latent-to-Action Policy Optimization (LAPO): A novel RL algorithm that jointly optimizes the latent reasoning process and action generation. By bridging reasoning and control, it improves representations for physical world modeling and enhances robustness in interactive environments. 2. Adaptive latent CoT mechanism: Allows the policy to dynamically adjust reasoning depth according to environmental complexity.

Results

LIBERO benchmark: LaST-R1 achieves a near-perfect 99.8% average success rate, requiring only a single supervised warm-up stage, and significantly surpasses state-of-the-art methods in both convergence speed and performance.
Real-world deployment: LAPO post-training yields improvements of up to 44% over the initial policy across four complex tasks, including single-arm and dual-arm setups.

---

*Auto-collected on 2026-05-03*