VLA-Precision: one hour of real-robot online RL reaches 98.3% success
Give a robot about 46 minutes to practice the real task and it reaches 98.3 percent average success. VLA-Precision starts from a pretrained vision-language-action model, puts it on the actual robot, and lets it improve through trial and error, using asymmetric co-bootstrapping for efficient real-world online reinforcement learning. Experiments cover nine precision chemistry tasks across four robot embodiments: attaching pipette tips, moving cuvettes, loading tube racks, inserting rubber stoppers and extinguishing an alcohol lamp. After an average 45.8 minutes of real-world training per task the model reaches 98.3 percent mean success, and its streaming training system improves throughput and computational efficiency by up to 10.9 times. Foundation models give a robot a strong starting point; an hour of practice then turns that general policy into a specialist.
