Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

On Agentic Manipulation Policies

Chen Chen, Omid Bagherieh, Naveen Nikamanth Appiah Balaji, Jesse T. Gonzalez, Negin Heravi, Fahad Islam, Xueting Li, Jean-Baptiste Regli, Kejia Ren, Petr Švec, David J. Yoon, Matthias Müller, Vladlen Koltun

November 2026

In a simulated RoboLab task, AMP1 picks up the banana and places it in the bowl.

TL;DR AMP1 lets a frontier AI model drive a robot arm directly by writing Python against a sensing, perception and motion SDK, with no demonstrations and no built-in grasp skills. It reaches 85.6% on RoboLab-120 against 49.2% for the best VLA, and succeeds in 104 of 108 real Franka trials against 49 for FLUX 3 Action and 30 for pi0.5.

What they do

Instead of training a VLA on demonstrations, AMP1 keeps a general-purpose frontier model (Gemini 4 Argon, Claude Opus 5.5 and others) in the control loop. The model writes code that looks at the cameras, commands the arm and revises its plan after each motion. One harness covers single-arm, bimanual and mobile manipulation in simulation and a real Franka, all zero-shot.

How it works

  • SDK in a persistent Python interpreter. Sensing (RGB, depth, calibration, robot state), perception that segments the fused point cloud into tracked objects the model can correct, and helpers to draw plans onto camera images.
  • Motion. The model sends timed end-effector waypoints and gripper commands; the harness solves IK and refuses unreachable or too-fast trajectories, with a reason. There are no grasp, place or push primitives, so the model writes them itself each episode.
  • Learning from experience (AMP1+). A persistent memory folder of notes and code that the agent rewrites after reflecting on binary rewards; model weights stay fixed.
  • Real robot. DROID setup (Franka Research 3, Robotiq 2F-85, two shoulder and one wrist camera), stereo depth, Opus 5.5 at high reasoning, plus SDK calls to recover from collision-reflex stops.

Results

  • RoboLab-120: 85.6% with Gemini 4 Argon and 82.8% with Opus 5.5, against 51.0% for the strongest learned policy on the leaderboard. With vague instructions AMP1 stays at 84.0% / 79.8%, while FLUX 3 Action drops from 42.9% to 28.4%.
  • CaP-Gym, zero-shot: 95.2% on LIBERO-PRO, 700/700 on robosuite, 48/50 on BEHAVIOR-1K.
  • Real world, 12 tasks with 9 trials each on three robot instances: 104/108 (94/108 without error recovery), against 49 for FLUX 3 Action and 30 for pi0.5.
  • The model matters most: Opus 4.5 reaches 35.2%. RGB-only input still gives 75.8% with Opus 5.5, but text descriptions instead of images give 36.8%. Practice with memory lifts RoboDojo bimanual success from 26.4% to 84.1%.

Limitations

The policy is slow: on the real tasks AMP1 averages 280 s to success, 211 s of it reasoning, against 44 s for FLUX 3 Action. The model cannot react while a trajectory runs, so evaluations are mostly quasi-static. Real-world failures come mainly from collisions, as motion planning is sensitive to camera calibration and uses simplified robot geometry.

Figures

From Figure 1: Zero-shot real-world episodes from a language instruction: (A) searching for hidden objects, (B) using...
From Figure 1: Zero-shot real-world episodes from a language instruction: (A) searching for hidden objects, (B) using a balance scale as a measuring instrument, (C) executing a winning chess strategy, (D) combining pushing and grasping where a direct grasp is infeasible, and (E) manipulating an articulated kitchen cabinet door. Instructions appear below each image row.
From Figure 6: The AMP1 harness. The generalist model controls the robot through Python code in a persistent interpre...
From Figure 6: The AMP1 harness. The generalist model controls the robot through Python code in a persistent interpreter and builds its own skills on top of an SDK with sensing, perception and motion functions. The memory (dashed) is used only when learning from experience.
From Figure 3: Emergent behaviors, (A to C) on the real robot and (D) in simulation, with quotes from the agent's thi...
From Figure 3: Emergent behaviors, (A to C) on the real robot and (D) in simulation, with quotes from the agent's thinking. (A) Moving an occluder to find a hidden target. (B) Sliding the plate against the bin and pushing an ungraspable bowl onto it. (C) Tipping an out-of-reach bottle toward itself with a wooden spoon. (D) Rolling a too-wide pomegranate up the crate wall. (E) Skills AMP1 writes for itself and how often each appears in 1,200 RoboLab episodes.
From Figure 2: Simulation benchmarks. (A) RoboLab-120 and the LIBERO-PRO, robosuite and BEHAVIOR-1K suites of CaP-Gym...
From Figure 2: Simulation benchmarks. (A) RoboLab-120 and the LIBERO-PRO, robosuite and BEHAVIOR-1K suites of CaP-Gym. (B) RoboLab-120 success rates with default, vague and specific instructions; AMP1 is highest. (C to E) Zero-shot CaP-Gym results: 95.2% on LIBERO-PRO, 100.0% on robosuite and 96.0% on BEHAVIOR-1K.
Abstract

Recent work on robot manipulation has focused on specialist policies trained on demonstrations. We show that generalist AI models can control manipulators more reliably than specialist policies such as vision-language-action models (VLAs) and world-action models (WAMs). We systematically analyze agentic manipulation policies (AMPs), in which a generalist AI model commands the manipulator directly, adapting its actions as the task unfolds. We present AMP1, a reference harness that runs unchanged across single-arm, bimanual, and mobile manipulation. AMP1 achieves state-of-the-art results on the RoboLab-120 and CaP-Gym benchmarks, with no demonstrations or task-specific training. On RoboLab-120, AMP1 reaches an 85.6% success rate zero-shot, compared to 49.2% and 42.9% for leading VLA and WAM policies. The agent succeeds zero-shot in 104 of 108 real-world trials spanning twelve tasks and three robot instances, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines. The basic harness supplies no grasping, pushing, or placing skills; the agent develops its own strategies, such as searching for hidden objects, using objects as tools, and calibrating its cameras. With the addition of a persistent memory store, the agent can also learn from experience: an augmented harness with a memory representation that is carried across episodes lowers the failure rate on RoboDojo bimanual manipulation tasks by 4.6x relative to zero-shot deployment. Our analysis suggests that AMPs are substantially more general and robust than specialist manipulation policies, providing a promising foundation for robot manipulation.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.