Skip to main content

AI-developed robot programs achieved 92.1% success across eight tasks

15 NOVEMBER 2026·2 MIN READ·3 SOURCES

The ACE-CaP-0 preprint reports that robot programs developed by Anthropic’s Claude Opus 5.5 coding agent succeeded in 221 of 240 real-world trials across eight manipulation tasks. The programs were frozen before evaluation and ran without a human or coding agent in the execution loop.

AI-developed robot programs achieved 92.1% success across eight tasks

Key takeaways · 4

  • 01

    For a benchmark reference, the frozen programs succeeded in 221 of 240 trials, or 92.1%, across eight real-world manipulation tasks.

  • 02

    The programs ran without a human or coding agent during evaluation, but still used runtime perception and control.

  • 03

    Development combined demonstrations on request with simulation on demand.

  • 04

    Treat the result as an evaluation of these tasks: the preprint says repeatability of the development process remains unestablished.

How ACE-CaP-0 develops programs

The method has a coding agent write and revise a robot program, request demonstrations when it lacks a strategy, and run simulated experiments to compare revisions.[3] Its development recipe combines demonstrations on request with simulation on demand.[1] The robot programs in the reported work were developed by Anthropic’s Claude Opus 5.5 coding agent.[2] In practice, this separates program development from evaluation: the agent can seek more input and test revisions while building a program, but the final program is assessed after it has been frozen.[3]

Results varied by task

The frozen programs achieved 92.1% success—221 successes in 240 trials—across eight real-world manipulation tasks.[3] Individual results ranged from 83.3% for bread packing to 100% for charger insertion.[1] The other reported task rates were 90% for bottle uncapping, flower insertion and shoe packing; 93.3% for cup nesting and parcel reorientation; and 96.7% for chicken plating.[1] Each program was evaluated for 30 trials without further updates.[3]

What ran during evaluation

Programs were frozen before evaluation and ran without a human or coding agent in the execution loop.[3] They still used runtime perception and control, so the result is not a test of a robot acting without those capabilities.[3] The evaluation used two Agilex PIPER X arms, one top-down camera and two wrist cameras.[3] Tasks included alignment, object reorientation, multi-object handling and bimanual coordination.[3] Mean execution times across tasks ranged from 0.97 to 2.27 minutes.[3]

Development results and an open question

The preprint reports that on-request guidance produced 86.7% final success, compared with 53.3% for another condition.[3] It also reports development time falling from 7.2 hours to 3.7 hours, and simulation-on-demand reaching 100% final success versus 86.7% for real-only development.[3] A separate comparison lists one development condition at 2.6 hours and another at 3.7 hours.[3] These findings describe the reported comparisons; the preprint says repeatability of the development process remains unestablished.[3]

For teams evaluating robotics, the useful distinction is between developing a program with agent assistance and executing a frozen program without an agent in the loop. The results can inform task-specific trials, but the stated repeatability limitation is a reason to validate the development process before relying on it.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 15 November 2026

    AI-developed robot programs achieved 92.1% success across eight tasks

Sources

AI fluency, one session a day, built for your work.