← November 16, 2026 briefing · added in the 20:00 KST update
ACE Robotics' ACE-CaP-0: AI agent writes robot control code and locks it in as reusable skills; 221 of 240 real-robot trials succeed (Pandaily)
Verified (fact)
According to Pandaily, in ACE Robotics' ACE-CaP-0 an AI agent writes robot control code. The finished code is frozen and used as a skill that can be reused in later tasks. CaP in the name stands for Code-as-Policies, an approach that uses the code itself as the robot's policy. In the researchers' own experiments, 221 of 240 real-robot trials succeeded, a 92.1% success rate. No external validation results have been released.
Needs fact-check (claim)
92.1% success rate in real-robot trials (based on the researchers' own tests).
Why it matters (interpretation)
Instead of a model directly outputting actions as in end-to-end VLAs, an LLM agent controls the robot through reviewable code. This is a case showing that the approach can achieve high success rates on real robots. It follows the same direction as ACE Robotics' previously covered 'Mundo'-based skill automation.
Confidence medium · official source pending
Sources
More about ACE Robotics
-
ACE Robotics automates end-to-end hotel laundry work with its Mundo world model and self-improving skills (report)
ACE Robotics combines its world model Mundo, a cloud planner, and on-robot skills to handle hotel laundry tasks from start to finish....
-
ACE Robotics and NTU S-Lab unveil 'Mundo,' a world model that builds physics-enabled 3D robot training scenes from 1–12 photos; code and weights due late November (report)
ACE Robotics and the S-Lab at Singapore's Nanyang Technological University (NTU) unveiled the world model 'Mundo.' It converts 1 to 12...
-
ACE Robotics releases ACE-Brain-0.6, a 9B model with Visual Grounding boosted by spatial data, claims it beats GPT-6 Astra on 6 of 8 evals (report)
According to Pandaily, ACE Robotics released ACE-Brain-0.6, a 9B model trained on about 1 million spatial data samples. However, the...