Model Releases
Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation
arXiv:2608.19490v1 Announce Type: cross Abstract: State-of-the-art vision-language-action (VLA) models such as pi_{0.5} exhibit strong semantic understanding, instruction following and task behavior.
arXiv:2608.19490v1 Announce Type: cross Abstract: State-of-the-art vision-language-action (VLA) models such as pi_{0.5} exhibit strong semantic understanding, instruction following and task behavior. However, when deployed on new robots, even minor mismatches in hardware configuration relative to pretraining can cause severe performance drops. Finetuning the VLA on in-domain expert data from the new embodiment improves performance on the expert task but leads to a loss in its original instruction following and behavioral priors. In this paper, we propose a self-supervised method that generates online interaction rollouts from the zero-shot VLA as additional training data for finetuning. Our experiments show this finetuning scheme yields strong multi-task policies that, on the target robot, (1) inherit prior tasks distilled from the zero-shot model, (2) enable generalist instruction following, while (3) learning new skills from expert data with improved sample efficiency. We demonstrate the success of our approach across test sets probing generalization on a real ALOHA robot and a new simulation benchmark in RoboTwin. Video results are available at https://self-supervised-control.pages.dev/
Related
- TEMPO: Semantic-Action Decoupled RL Post-Training for Vision-Language-Action Models
- AEGIS: Anchor-Enforced Gradient Isolation for Knowledge-Preserving Vision-Language-Action Fine-Tuning
- ReFineVLA: Multimodal Reasoning-Aware Generalist Robotic Policies via Teacher-Guided Fine-Tuning
- S^2-VLA: State-Space Guided Vision-Language-Action Models for Long-Horizon Manipulation
Source: arXiv cs.CV | 2026-08-21