The Low-Rank Structure of VLA Reinforcement Learning

Despite the impressive visuomotor capabilities enabled by Vision-Language-Action (VLA) models, their performance often degrades on challenging and out-of-domain tasks. Recent test-time steering and scaling methods improve performance without extensive data collection and retraining, but action samples often remain concentrated around similar behaviors and therefore inherit correlated failure modes. Moreover, existing methods apply the same intervention strategy at every timestep, regardless of whether the base policy is already likely to succeed. To address these limitations, we introduce $RL^2$, an adaptive inference-time steering framework that leverages Reinforcement Learning on VLA Latents. First, we train a lightweight offline RL policy conditioned on expressive latents extracted from the VLA action expert and compose its flow velocity with that of the frozen VLA during inference. This compositional steering strategy combines the behavioral priors of large-scale imitation learning with the action diversity induced by offline RL beyond dominant demonstration modes. We further discover that inference-time steering follows fundamentally different scaling laws under success and failure states, revealing that action diversity is most beneficial when the base VLA is likely to fail, but can unnecessarily perturb already-accurate actions when success is likely. Building on this insight, $RL^2$ activates compositional steering only when failure is predicted. Across the SIMPLER and PolaRiS benchmarks, $RL^2$ improves success rates by up to +17.3% in out-of-domain settings, while ablations and scaling studies demonstrate the importance of latent representations and RL training. Finally, real-world experiments demonstrate that these gains transfer beyond simulation, establishing $RL^2$ as a practical and modular steering framework for VLA deployment.

Reinforcement learning is increasingly used to fine-tune vision-language-action (VLA) models, but robot interaction is expensive and learning becomes highly sample inefficient when successful rollouts are rare. When reward is assigned only for completing the commanded task, a failed rollout is treated as having no value even if it successfully executes behaviors relevant to that task. A robot that fails to place the correct object in a bowl may still move that object toward the bowl or place a different object inside it, demonstrating objects and actions that can be reused to solve the target task. These behaviors define auxiliary tasks that the policy can already solve, providing useful learning signals even before it can solve the harder target task. We introduce $\textit{Learning from Hindsight (LfH)}$, which turns such failures into additional learning signals. Using a pretrained vision-language model, LfH relabels failed rollouts with the behaviors they actually accomplish and trains the policy jointly on the commanded task and these auxiliary tasks. On out-of-distribution LIBERO-PRO manipulation tasks, LfH matches the final performance of GRPO with approximately $5\times$ fewer rollouts and improves sample efficiency across multiple VLA backbones. On a physical Franka robot, LfH raises success from $0\%$ to $56\%$ within 160 training rollouts, while GRPO reaches $22\%$.

Reinforcement learning (RL) fine-tuning has shown promise for Vision-Language-Action (VLA) models in robotic manipulation, but deployment-time visual shifts pose practical challenges. A key difficulty is that standard task rewards supervise task success, but offer limited guidance on whether a visual change is task-irrelevant or changes the behavior required for manipulation. We propose PAIR-VLA (Paired Action Invariance & Sensitivity for Visually Robust VLA), an RL fine-tuning framework to address this difficulty by adding two auxiliary objectives over paired visual variants during PPO optimization: an invariance term that reduces the discrepancy between action distributions for a task-preserving pair (e.g., different distractors), and a sensitivity objective that encourages separable action distributions for a task-altering pair (e.g., target object in a different pose). Together, these objectives turn visual variants from mere observation diversity into behavior-level guidance on policy responses during RL fine-tuning. We evaluate on ManiSkill3 across two representative VLA architectures, OpenVLA and $π_{0.5}$, under diverse out-of-distribution visual shifts including unseen distractors, texture changes, target object pose variation, viewpoint shifts, and lighting changes. Our method consistently improves over standard PPO, achieving average improvements of 16.62% on $π_{0.5}$ and 9.10% on OpenVLA. Notably, ablations further show generalization across visual shifts: invariance guidance learned from distractor and texture variants transfers to target-pose and lighting shifts, while adding sensitivity guidance on target-pose variants further improves robustness to nuisance shifts, highlighting the broader transferability of behavior-level RL guidance.