Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

ACE-CaP-0: Real-World Code-as-Policies in the Era of Experience

Changxu Wei, Yunpeng Qing, Guozheng Ma, Bowen Liu, Zexin He, Yilun Kong, Haoming Gu, Jingyuan Liu, Tianyu Zhu, Jianfei Wang, Junqiang Kou, Huinan Mo, Hao Lv, Pengbin Li, Yuesong Wang, Bo Li, Huan Zhou, Shunyu Liu, Chunxiao Liu, Bixuan Jin, Shaopeng Yin, Fei Wang, Zhi Hou, Jiaxing Huang, Changqing Zou, Dacheng Tao, Xiaogang Wang

November 2026

TL;DR A coding agent writes robot programs, asking for kinesthetic demonstrations and running simulated experiments only when it judges they will help, then freezes the code. On eight real manipulation tasks with a two-arm robot, the frozen programs succeed in 221 of 240 trials (92.1%) with no human or agent in the loop.

What they do

With Code as Policies, what a coding agent works out while developing a skill is kept in a program that later runs without model calls. Physical trials are expensive, so the open question is how the agent should gather experience. On charger insertion they compare forms and timing of human guidance and ways of using simulation, then combine the best options into a recipe, ACE-CaP-0, and test it on eight real tasks.

How it works

  • Development loop. The coding agent (Claude Opus 5.5) writes, runs, inspects and revises a program that calls the perception and control interfaces of two Agilex PIPERX arms with one top-down and two wrist cameras. Each task gets at most ten human-assisted scene resets.
  • Demonstrations on request. When the agent asks, a person guides the arms kinesthetically; the agent receives the video, arm trajectories and key joint poses.
  • Simulation on demand. The agent can build or modify a simulator to compare candidate revisions on a chosen task segment and initial state, then checks the result on the robot.
  • Frozen execution. The final program is frozen and runs with runtime perception and control, with no human or agent in the loop.

Results

  • Guidance form on charger insertion (30 trials each): robot demonstrations 86.7%, human video from the robot’s viewpoint 73.3%, text only 60%.
  • Guidance on request reaches 86.7% after 3.7 h of development; the same guidance given upfront reaches 53.3% after 7.2 h.
  • Simulation on demand reaches 100% in 2.6 h, at more output tokens (1.09M vs 0.73M for real-only). Refining fully in simulation before transfer scores 0%: the grasp fails because simulated images and depth were cleaner than the real sensor’s.
  • Across eight tasks the frozen programs succeed in 221/240 trials, from 83.3% (bread packing) to 100% (charger insertion), with mean execution times of 0.97-2.27 min.

Limitations

Each condition in the investigation is one development run on one task, so consistency across development runs is not established. Cost accounting leaves out demonstration effort, and reusing the programs under new conditions is untested.

Figures

From Figure 1: A coding agent refines policies on the robot, requesting demonstrations and simulation as needed (left...
From Figure 1: A coding agent refines policies on the robot, requesting demonstrations and simulation as needed (left), and the eight tasks with the success rate and mean execution time of each frozen policy over 30 trials (right).
From Figure 5: (a) The coding agent revises an executable policy from real-world feedback, requesting kinesthetic dem...
From Figure 5: (a) The coding agent revises an executable policy from real-world feedback, requesting kinesthetic demonstrations (video, arm trajectories, key joint poses) or testing candidate code in simulation for a chosen task segment. (b) The final policy is frozen and runs without a human or coding agent in the loop.
From Figure 4: Three ways to use simulation (real-only, ten simulation rounds before transfer, simulation on demand d...
From Figure 4: Three ways to use simulation (real-only, ten simulation rounds before transfer, simulation on demand during real-world development) and, for charger insertion, each frozen program's success over 30 real trials, development time and output tokens.
From Figure 7: The eight real-world tasks and the success rate of each frozen code policy over 30 evaluation trials.
From Figure 7: The eight real-world tasks and the success rate of each frozen code policy over 30 evaluation trials.
Abstract

Recent advances in frontier models have pushed the boundaries of agentic robot control, enabling agents to perform complex robotic tasks through reasoning and interaction. However, repeated calls to these models during execution can substantially increase task completion time. Code as Policies (CaP) offers a route for coding agents to acquire physical skills through experience and retain them in executable programs, allowing the reasoning invested in development to be reused across runs. Developing reliable code policies while limiting the cost of physical experimentation remains a central challenge. We investigate human guidance as a source of priors for feasible actions and simulation as an auxiliary environment for exploring and testing candidate policies. Our findings favor robot demonstrations available on request and simulation used selectively during real-world policy refinement. These findings inform ACE-CaP-0, an agentic recipe that couples policy revision with experience acquisition through Demonstrations on Request and Simulation on Demand. Across eight real-world manipulation tasks spanning alignment, multi-object handling, and bimanual coordination, the resulting frozen programs achieve 92.1% success (221/240 trials). Mean execution times across tasks range from 0.97 to 2.27 minutes. The programs execute with runtime perception and control, without a human or coding agent in the loop. These results highlight CaP's potential to turn experience into reliable, reusable physical skills.

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.