APL

AMP1

Apple, United States 路 November 2026

Agentic policy 路 Closed weights 路 Also written Agentic Manipulation Policies, On Agentic Manipulation Policies, AMP, ml-amp

Parameters
not published
Not applicable.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
No weights exist: AMP1 is a harness around third-party frontier models.

What it is

AMP1 is Apple's reference harness for agentic manipulation policies: instead of a trained robot model, a general purpose frontier model such as Claude Opus 5.5 or Gemini 4 Argon controls the arm by writing and running Python code over several model calls. The harness gives it sensing (RGB-D images, calibration, a fused point cloud), a scene model it can correct and motion commands as timed end-effector waypoints, but no grasping, pushing or placing skills. Without demonstrations or task-specific training it leads NVIDIA's RoboLab-120 leaderboard at 86.0 percent, against 49.2 percent for the best VLA and 46.2 percent for the best world action model on that board, and the paper reports 104 successes in 108 real-world trials. There are no weights to download: the code is public, and every episode needs access to a frontier model. The authors themselves call such policies not yet practical, because the robot spends much of each episode waiting for the model to reason.

Architecture

Backbone
None of its own. The harness drives a frontier model through its API: Claude through the Anthropic Messages API, Gemini through generateContent, and open-weights models through an OpenAI-compatible server (sglang or vLLM).
Action head
No learned action head. The model writes Python against the harness SDK and commands the arm through timed end-effector waypoints and gripper commands, which the harness checks for reachability, speed, joint limits and collisions before it executes them. A single turn can command a whole trajectory, and images enter the model's context only when its code displays them.
Parameters
Not applicable. AMP1 has no parameters of its own; size and cost are those of the frontier model it calls.
Pretraining data
None. No demonstrations and no task-specific training; the harness relies on what the frontier model already knows. An augmented harness adds a memory store that is carried across episodes, so the agent can learn from practice.
Embodiments
Franka with Robotiq 2F-85 gripper (DROID setup, RoboLab-120), Single-arm, bimanual and mobile manipulators in simulation (CaP-Gym: LIBERO-PRO, robosuite, BEHAVIOR-1K), Bimanual arms (RoboDojo), Three real robot instances (paper)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

The policy itself needs no GPU, the model runs behind an API. The paper's simulation results were run on NVIDIA H100 80 GB GPUs for the simulators. With Claude Opus 5.5 a RoboLab-120 episode costs 1.39 US dollars on average at list price, according to the paper.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • CaP-Gym robosuite, single-arm and bimanual, zero-shot
    100%
    success rateAuthors' own evaluation. HarnessPAI reaches 96.3 percent after practice, according to the paper.source
  • CaP-Gym BEHAVIOR-1K, mobile manipulation, zero-shot
    96%
    success rateAuthors' own evaluation.source
  • CaP-Gym LIBERO-PRO, zero-shot
    95.2%
    success rateAuthors' own evaluation. HarnessPAI reaches 96.5 percent after practice, according to the paper.source
  • RoboLab-120 overall, 25 episodes per task
    86%
    success rate2,579 of 3,000 episodes on the public leaderboard, zero-shot. Listed as a submission without the organizers' verified mark as of 2026-11-22. The paper's abstract reports 85.6 percent.source
  • RoboDojo bimanual manipulation, augmented harness with memory, after practice
    84.1%
    success rateAuthors' own evaluation. The memory store lowers the failure rate 4.6 times relative to zero-shot deployment, according to the abstract.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real robot, twelve tasks on three robot instances (AMP1 paper) zero-shot
    96.3%
    success rateAuthors' own evaluation: 104 of 108 trials. In the same trials FLUX 3 Action succeeded in 49 and the VLA baseline in 30.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

104 successes in 108 zero-shot trials across twelve tasks and three robot instances, against 49 for FLUX 3 Action and 30 for the VLA baseline in the same trials. The paper documents strategies the agent develops on its own, such as searching for hidden objects, using objects as tools and calibrating its cameras.

Fine tuning it yourself

Nothing to fine tune. You choose the frontier model and its reasoning effort, and the harness runs unchanged on a new robot. The repository reproduces the paper's results on RoboLab-120, CaP-Gym and RoboDojo; it needs Linux with NVIDIA GPUs for the simulators and an API key for the model, or a self-hosted open-weights model behind an OpenAI-compatible server.

Where it helps, where it does not

Strengths

  • Highest success rate on RoboLab-120: 86.0 percent zero-shot, against 49.2 percent for the best VLA and 46.2 percent for the best world action model on the same board.
  • No demonstrations, no training and no built-in manipulation skills: the same harness runs single-arm, bimanual and mobile manipulation.
  • Strong on real hardware in the authors' trials: 104 of 108, with 15 times fewer failures than FLUX 3 Action and 20 times fewer than the VLA baseline.
  • Cheaper than earlier agent approaches: about 1.39 US dollars per RoboLab-120 episode with Claude Opus 5.5, about 26 times below the 36.40 US dollars reported for Astranomy with GPT-6 Astra.

Limits

  • Not a model: there are no weights, and every episode depends on a frontier model, usually a paid one behind an API.
  • Slow. The authors write that agentic policies are not yet practical, because the robot spends much of each episode waiting for the model to reason.
  • Results move with the model and its reasoning effort: the paper's cost analysis lists 82.8 percent on RoboLab-120 with Claude Opus 5.5.
  • The code is under Apple's own licence, and the RoboLab-120 entry does not carry the organizers' verified mark.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-11-22.

Try a policy on a real arm

A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.