ACE Robotics and academic collaborators describe ACE‑CaP‑0 in a preprint submitted to arXiv on November 12, 2026. An agent develops a robot program that is subsequently frozen and executes without a coding agent or human in the loop. The authors report 221 successful trials out of 240 across eight real-world manipulation tasks, or 92.1%. These are results from their setup, not independent SIExpo testing.

Develop once, execute repeatedly

The project page describes two Agilex PIPER X arms, cameras and Claude Opus 5.5 as the coding agent. Each frozen program receives 30 evaluation trials. Execution time is distinguished from the time and tokens spent during development. That separation is central to the practical interest of the work.

Our editorial reading is that retaining code changes where expensive reasoning takes place. Instead of repeating the full programming process during every subsequent operation, the developed logic is reused. This does not imply operation without observations: fixed rules can still use current sensor information. A frozen program is not synonymous with identical movement regardless of the scene.

When to request assistance

The researchers examine demonstrations requested when needed and simulation used for a specific purpose. The project explicitly cautions that each development condition in the comparison was run once on one task. The 30 subsequent trials measure the reliability of the resulting program; they do not establish repeatability of the complete process that created it.

That is an important statistical boundary. A well-performing final program still leaves the question of how often another independent development run would produce an equally good one. For a production decision, these questions have different implications: one concerns operation after setup, while the other concerns the cost and predictability of setup itself.

Another useful boundary lies between a digital twin and actual contact with objects. Simulation can help examine a geometric question, but the value of that check depends on which feature of the real scene it represents. The conclusion is not that simulation is unhelpful; it is that its intended question must remain explicit.

Reading the percentage

The 92.1% figure applies to these eight tasks, this equipment and this evaluation protocol. It does not establish equivalent behavior with different objects, a new camera or an unfamiliar robot. The programs offer an interesting route for retaining experience, while the scope of that experience remains part of the claim.

For a future assessment, it would be useful to keep program-development records alongside execution results. That would make it possible to examine which intervention produced a change and whether it remained useful across later trials. This is an editorial suggestion for evaluating the approach, not an additional result reported by the study.

The preprint is older than the usual daily window and was selected as valuable uncovered material within seven days. We do not present it as a November 16 release or an established industrial system. SIExpo read the original paper and project page without reproducing the physical experiments.

Editorial verification: 2026-11-16T17:16:38.474339+00:00.