JetsonsAge

Explore the frontier.

Search published reporting, intelligence and briefs.

Explore JetsonsAge

Robots Keep AI Developed Skills as Reusable Programs

Researchers have developed robot skills as reusable software, concentrating an AI coding agent’s work in the development process.

Published November 16, 2026Briefs / Robotics / Autonomous Systems

Researchers have developed robot skills as reusable software, concentrating an AI coding agent’s work in the development process. Once the programs are finished, robots can use them repeatedly while responding to cameras and other control inputs.

The ACE-CaP-0 team reports 221 successful trials out of 240 across eight manipulation tasks. The work comes from ACE Robotics and collaborators at the Hong Kong Polytechnic University, Zhejiang University and Nanyang Technological University.

For a business considering robots, the attraction is a skill that survives as an inspectable program. Engineers could keep the code, examine how it handles a task and evaluate a revised version before using it again.

The November 12 preprint develops an approach with a research history stretching back several years: asking language models to write robot-control programs.

How experience becomes a program

Google researchers’ 2022 Code as Policies project showed language models writing code from natural-language instructions. Those programs could combine perception with ordinary programming structures and calls to robot-control functions. The project demonstrated manipulation, drawing and mobile-robot tasks.

In robotics, a policy is the rule that connects a situation to an action. A code-based policy can contain branches or repeated steps, with a camera supplying information as the program runs. A saved program can therefore react to a changing scene.

ACE-CaP-0 adds a development loop in which a coding agent requests experience. A person can guide the robot through a demonstration, and simulation can help investigate candidate actions. The agent uses the results to revise its code.

The team then froze the programs for evaluation, giving each task 30 physical trials. Those tests produced the reported 92.1 percent success rate, including tasks such as inserting a charger and opening a bottle. Runtime perception and control remained active.

The reusable part of robot learning

The method’s contribution is the path from demonstrations and trials into executable instructions. It gives developers another way to retain experience alongside approaches that encode skills in a trained model’s parameters.

Carrying these programs into a workplace requires handling varied objects, starting positions and interruptions. Development also includes demonstrations, physical testing and the work of diagnosing errors.

Sources

Sources · 3
More briefs