Self-Improving Vision-Language-Action Models with Data Generation via Residual RL

Supervised fine-tuning (SFT) has become the de facto post-training strategy for large vision-language-action (VLA) models, but its reliance on costly human demonstrations limits scalability and generalization. We propose Probe, Learn, Distill (PLD), a three-stage plug-and-play framework that improves VLAs through residual reinforcement learning (RL) and distribution-aware data collection: train residual specialists to handle failure cases, collect diverse data via hybrid policy rollout, and distill the resulting trajectories back into the generalist via SFT. We evaluate PLD across diverse settings: it achieves a near-saturated 99% task success rate on the LIBERO benchmark, delivers over 50% performance gains in SimplerEnv, and demonstrates a 100% success rate on real-world Franka arm and YAM arm dexterous manipulation tasks. We further provide ablations showing that residual policy probing and distribution-aware replay are key to collecting deployment-aligned data that improves VLAs’ capabilities on both seen and unseen tasks. Our results demonstrate that RL-generated, policy-aligned data can surpass teleoperation-only demonstrations, offering a scalable path toward self-improving VLA models. More results can be found on website.