ACE Robotics' ACE-CaP-0 lets AI agents write reusable robot code, reporting 92.1% success
ACE Robotics has released ACE-CaP-0, a research recipe in which a general-purpose AI agent writes, tests and refines robot control programs, then freezes them into reusable skills that run without the agent. The work reports 92.1% success. The paper was posted to arXiv on November 12.
ACE Robotics has released ACE-CaP-0, a research recipe in which a general-purpose AI agent writes, tests and refines robot control programs, then freezes them into reusable skills that run without the agent. The work reports a 92.1% success rate, according to the report by Pandaily.
ACE Robotics is an embodied AI company chaired by SenseTime co-founder Wang Xiaogang. The paper was posted to arXiv on November 12 with an accompanying project page, and the company announced the work on November 13.
How it works
ACE-CaP-0 builds on Code as Policies, an approach that expresses robot behavior as executable code instead of calling a large model at every step. The authors argue that repeated model calls slow a robot down, while a finished program reuses the reasoning that went into writing it.
In ACE-CaP-0, the agent writes the control program, tests it and refines it. Once the program is finished, it is frozen into a reusable skill. The skill then runs without the agent in the loop.
The authors identify the central difficulty as developing reliable code without too many costly physical trials. In ACE-CaP-0, the agent decides for itself when to seek help.
Key facts
- Developer: ACE Robotics
- Release: ACE-CaP-0, described as a research recipe
- Reported success rate: 92.1%
- Paper: posted to arXiv on November 12, with a project page
- Company announcement: November 13
- Pricing, context window, parameter count: not applicable or not disclosed
What this means
ACE-CaP-0 targets a specific cost in robot control: the latency of calling a large model at each step. By moving the agent's work to the authoring stage and freezing the result into a program, the design keeps the agent out of the runtime loop, which is the authors' stated reason for the approach. The 92.1% figure is the company's reported result, and readers evaluating it should consult the arXiv paper and project page for the task setup and baselines behind it.
Related Articles
Salesforce's SFR-AutoR&D agents built SFR-Embedding v4 in about 80 hours
Salesforce introduced SFR-AutoR&D on November 11, 2026. The company says its agents produced SFR-Embedding v4 in about 80 hours, compared with the several weeks that human experimentation often requires.
Microsoft researchers backdoor LLM agents by reprogramming fewer than 10 dormant neurons, with no training
Microsoft researchers describe Lazarus, a method that reprograms fewer than 10 dormant neurons in an LLM to backdoor an agent built on it, with no training required. The modified model passes standard evals until a trigger appears. Microsoft says provenance and weight integrity are the key defenses.
NVIDIA recipe: RL post-training lifts Nemotron 3.5 Super VL accuracy a further 17% on drug-metabolism task
NVIDIA has published a recipe for post-training Nemotron 3.5 Super VL on scientific tasks using NeMo Gym and NeMo RL. In a drug-metabolism example, supervised fine-tuning improved accuracy by 12% and RL post-training lifted it a further 17% at a 16,384-token output budget.
Anthropic: Selective notes beat full logs for Claude agent memory, with 'Dreaming' fixing errors
Anthropic published research on how Claude retains knowledge across sessions using note files. Its tests show selective notes outperform exhaustive archives, and a post-session process called Dreaming can correct faulty notes. All results come from Anthropic's own tests.
Comments
Loading...