Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation
Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation
This paper addresses performance drops in vision-language-action (VLA) models like pi0.5 when deployed on new robots due to hardware mismatches. The authors propose a self-supervised method that generates online interaction rollouts from a zero-shot VLA to create additional training data for fine-tuning. Experiments show this approach yields strong multi-task manipulation policies that inherit pretrained tasks, improve instruction following beyond expert-only fine-tuning, and mitigate forgetting of held-out tasks. Results are demonstrated on a real ALOHA robot and a new simulation benchmark in RoboTwin.
The authors propose a self-supervised method using online interaction rollouts from a zero-shot VLA to fine-tune models for multi-task manipulation on new robots, demonstrating improved generalization and reduced forgetting compared to expert-only fine-tuning.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org