A generalist AI model that watches a robot’s surroundings, writes code and revises its actions during a task achieved higher completion rates than specialist controllers in tests reported by Apple-affiliated researchers. Their November 6, 2026 arXiv preprint, “On Agentic Manipulation Policies,” presents AMP1 as a promising research approach, although substantial reasoning delays constrain practical use. Peer review has not been verified, and no independent replication is supplied.
AMP1 is a software harness supplying camera observations, code execution and interfaces for commanding robot motion. Its basic configuration includes no predefined grasping, pushing or placing skills: the model chooses how to manipulate objects and responds to new observations. Unlike approaches that generate a program before deployment, the generalist model remains in the control loop. Here, “zero-shot” means no task-specific demonstrations, training or practice rollouts during evaluation; it does not establish that benchmark material was absent from model pretraining.
On RoboLab-120, AMP1 paired with Gemini 4 Argon completed 85.6% of 1,200 episodes under default instructions, with a reported 95% confidence interval of 83.5–87.5%. Leaderboard results for the strongest vision-language-action and world-action models were 49.2% and 42.9%, respectively—gaps of 36.4 and 42.7 percentage points. Across three CaP-Gym suites, AMP1 reported 95.2% success over 3,000 LIBERO-PRO episodes, 700 successes in 700 robosuite episodes and 48 in 50 episodes on two BEHAVIOR-1K tasks. Some competing agents practiced on evaluated tasks, complicating comparisons.
Physical tests used a Franka Research 3 arm, a Robotiq 2F-85 gripper and three cameras. Each system attempted 12 author-designed tasks across three robot instances, with nine trials per task. AMP1 with Opus 5.5 succeeded in 104 of 108 attempts, or 96.3%, versus 49 for FLUX 3 Action and 30 for π0.5. Those counts imply approximately 15-fold and 20-fold lower failure rates, respectively.
That comparison includes a consequential difference: AMP1 could recover from robot controller errors, while baseline trials ended when such errors interrupted motion. Without recovery, AMP1 completed 94 of 108 trials, or 87.0%. It was also slow. Successful AMP1 trials averaged 280 seconds, including 211 seconds of reasoning, versus 44 seconds for FLUX and 62 seconds for π0.5. These averages cover different subsets of successful trials and are therefore not matched speed comparisons.
Separate experiments added persistent memory containing notes or code while leaving model weights fixed. On 42 simulated bimanual RoboDojo tasks, practice raised reported success on unseen layouts from 26.4% to 84.1%; a single continual-learning pass across tasks reached 57.6%. These are experience-based results, distinct from zero-shot performance.
Selected episodes showed the agent moving obstructions to find hidden objects, using tools and rearranging objects to make grasping possible. They illustrate strategies rather than establish their general reliability. Controlled experiments also found that model choice strongly affected performance and that stronger models could compensate for missing sensory information through additional computation.
The evidence supports higher success under the tested protocols, with important limits. Physical testing covered one arm-and-gripper platform and 12 tasks, while models, harnesses and training histories differed across systems. Repeated episodes share tasks and environments; whether the reported confidence interval accounts for that dependence is unclear. Detailed methods and supplements referenced by the paper are not fully available in the supplied material. The findings do not establish universal superiority or deployment readiness.