Luma AI (Benjamin Pikus, Daniel Saunders, Mircea Mironenco and co-authors) · Sim-trained robot agent

Zero-real Agent Policy (ZAP)

Zero-real Agent Policy (ZAP) is a robot policy from Luma AI that teaches a pretrained 27B vision-language model motion control entirely in simulation, using no real-world robot or teleoperation data. The model acts through a minimal harness that sets end-effector poses and gripper openings, with no grasp macros or motion primitives, and the same harness runs in simulation and on hardware. It is trained on simple simulated pick-and-place tasks, learning from its own successes first by fine-tuning and then with GRPO, with no frontier teacher to distill from. On 10 real-world tasks selected from RoboLab-120 (30 trials), ZAP reaches 97% success, level with GPT-6 Astra in the same harness, while the base model rarely succeeds. ZAP averages 146 s per task against 397 s for GPT-6 Astra, and $0.29 against $4.87 per task. The authors also show zero-shot transfer to a real bimanual robot on unseen tasks such as a mid-air test-tube handoff, drawing on a tic-tac-toe board and finding a mate-in-one in chess. The text does not state a licence, code release or version.