Real-robot self-improvement on OpenMANIPULATOR-X
Fine-tuning π0.5 on its own rollouts with a low-cost arm (in progress)
An ongoing study of what a vision-language-action model gains and loses when it keeps learning after deployment, run on a real OpenMANIPULATOR-X arm rather than in simulation. The policy is π0.5, fine-tuned on a case-to-mug pick-and-place task, then fine-tuned again on its own self-generated rollouts.
Artifacts on Hugging Face
Videos
Videos coming soon.