A Guth Labs publication ↗
Back to News

ACE-CaP-0 turns robot experimentation into reusable code, not constant model calls

ACE-CaP-0 turns robot experimentation into reusable code, not constant model calls

A robot need not ask a large model what to do at every step. ACE-CaP-0 explores a different division of labor: use a coding agent to develop a skill, then run the resulting program with perception and control.

In their November 12 preprint, the researchers describe a development loop that can request robot demonstrations, use simulation selectively and revise code from physical feedback. Frozen programs succeeded in 221 of 240 trials across eight manipulation tasks. Mean execution times ranged from 0.97 to 2.27 minutes. The coding agent and human were absent from that execution loop. Paper

That is not a claim that skill development takes a minute. The paper separates development from evaluation, and allows human-assisted scene resets during development. Its exploratory comparisons use one development run per condition. These are author-reported results, not independent replication or a demonstration of general household competence. Full paper

Guth analysis

The useful question is where reasoning should happen. Moving it into skill development can make repeated execution easier to inspect and budget. It also creates a different failure mode: a frozen program may be efficient inside its tested conditions but brittle outside them. Buyers should ask for the operating envelope, a failure detector and a recovery procedure before treating a successful demonstration as a deployable worker.

Evidence limits

This is newly covered research from November 12, not breaking news. The authors' project page was inaccessible in this pass; no code-release or licensing claim is made. No independent reproduction was located.

Sources and citations

Revision history

Published .