Training pi0.5 on OpenArm: from teleop data to a 100% pick-and-place policy

I've been working on a robotics contract at APTO, a Japanese startup, for a while now. I had a chance to prototype pi0.5 models with OpenArm, and it was a very educational experience. pi0.5 is Physical Intelligence's 3.6B-parameter VLA (Vision-Language-Action model), and we fine-tuned it on the open-source bimanual OpenArm robot with LeRobot, going from teleop data collection all the way to a policy that hits 100% success on a pick-and-place task.

I wrote a blog post about the whole process on APTO's tech blog. The post covers the full pipeline: the teleop recording cell, the training recipe, real-time chunking inference, robustness tests under sensor loss, and three silent pitfalls that never raised an error but cost us the most time.

Read the full article here:

Training pi0.5 on OpenArm: from teleop data to a 100% pick-and-place policy →