Researchers led by Artur Kuramshin have studied how best to adapt vision-language-action models to multi-fingered robot hands, according to a paper on arXiv. Most manipulation data comes from parallel-jaw grippers, and demonstrations are harder to collect by teleoperating hands with several fingers.
Using the π0.5 model and a LEAP hand, the team tested five recipes that combine supervised fine-tuning with diffusion-steering reinforcement learning on two simulated tasks with only a few demonstrations. Success after fine-tuning was a poor guide to success after steering, so the best fine-tuned checkpoint was seldom the best one to steer.
Freezing the vision-language model and training only the action expert gave the policy that was easiest to steer, also for object placements outside the training data, while low-rank adaptation was the hardest. The authors recommend that recipe and advise against choosing checkpoints by their success after fine-tuning. The paper was accepted at a CoRL 2026 workshop on dexterous manipulation.



