Apple AMP1 robotic arm succeeded 104 out of 108 trials

Apple researchers reported that their AMP1 robot-arm control system succeeded in 104 of 108 real-world trials, far outperforming the purpose-built FLUX 3 Action model (49 successes) and π0.5 (30 successes), according to a paper preprint posted to arXiv on the 6th. The system, described in “On Agentic Manipulation Policies,” has a general-purpose AI model write and execute Python code that moves the robot arm, feeding camera images, depth data and robot-state information into the model and re-planning each next action based on newly observed information during a task. The code runs in a continuously operating Python interpreter that can check robot state, distinguishing the approach from conventional end-to-end control policies. The result matters because it suggests general AI models paired with code execution can outperform specialized manipulation models on physical tasks, a potential shift in how robotics control stacks are built. What to watch next is whether the arXiv findings hold up under peer review and whether the approach scales beyond the lab to broader manipulation tasks and hardware.