A general model commands the arm directly: AMP1 hits 85.6% zero-shot on RoboLab-120, ahead of VLA and WAM policies
Synopsis
Robot manipulation has relied on specialist policies trained on demonstrations, such as vision-language-action models (VLAs) and world-action models (WAMs). AMP1 keeps a generalist AI model in the control loop and has it command arm motion through code, reaching 85.6% zero-shot success on RoboLab-120 against 49.2% and 42.9% for leading VLA and WAM policies. It succeeded in 104 of 108 real-world trials across twelve tasks and three robot instances, with no demonstrations or task-specific training.
Interpretation
A generalist AI model can command a manipulator directly and revise its actions as a task unfolds, and this route outperforms VLA and WAM policies trained specifically for robot control on both simulated benchmarks and real hardware. Earlier uses of generalist models mostly orchestrated existing low-level skills, which bounded their control authority, or synthesized a control program offline, so the model could not reconsider the overall strategy during execution. On RoboLab-120, AMP1 with Gemini 4 Argon reaches 85.6% zero-shot success, while the strongest VLA and WAM entries reach 49.2% and 42.9%; on real hardware it succeeds in 104 of 108 trials, against 49 for FLUX 3 Action and 30 for another VLA baseline.
AMP1 is a reference harness that gives the model visual observations, code execution and direct control of manipulator motion, and the same harness runs unchanged across single-arm, bimanual and mobile manipulation. The harness supplies no grasping, pushing or placing skills; the model writes its own geometry and control code each episode, from choosing a grasp to measuring a container. The same harness is used on RoboLab-120, LIBERO-PRO, robosuite and BEHAVIOR-1K, with only the robot description and camera configuration changing; real trials use a 7-DoF Franka Research 3 arm with a Robotiq 2F-85 gripper and a three-camera system.
Perspective
The demonstrated scope is zero-shot evaluation on four simulation benchmarks plus twelve real-world tasks on one 7-DoF Franka Research 3 arm with a Robotiq 2F-85 gripper and a three-camera system, across three platform instances and nine trials per task; the harness stays the same across single-arm, bimanual and mobile manipulation in simulation, while real experiments involve only a single-arm platform.
The authors state that AMP1 is not yet practical: the robot spends much of each episode waiting for the model to reason, and the model cannot reconsider a trajectory while it executes; evaluations largely concern quasi-static scenes, whereas a moving target or a slipping object may require a response before the model's next turn.