Gasgoo Munich- ACE Robotics recently released its latest research findings, ACE-CaP-0. The study investigates how general-purpose models can autonomously develop, verify, and optimize robotic skills by actively acquiring experience—cutting execution times for precision tasks from hours to just minutes while significantly boosting success rates. Image source: ACE Robotics ACE-CaP-0 innovatively fuses on-demand demonstration with on-demand simulation mechanisms. During skill development, the AI agent autonomously determines when to request human guidance and when to run simulations, continuously testing and refining code strategies based on real-world feedback. Once verified, these operational methods are solidified into reusable, executable robotic skills—effectively decoupling high-cost intelligent exploration from high-frequency execution. Across eight real-world precision tasks—including plugging in a charger, packing bread, and opening bottle caps—code strategies developed by ACE-CaP-0 slashed average execution times from hours to under a minute. In 240 independent real-world trials, the strategy succeeded 221 times, achieving an overall success rate of 92.1%. Notably, seven tasks saw success rates of at least 90%, with the charger-plugging task hitting a perfect 100%. Take the charger-plugging task as a case in point: a general-purpose model required roughly an hour of trial, error, and processing time to dynamically adjust its strategy. By contrast, ACE-CaP-0 distilled effective methods into reusable skills through iterative optimization, enabling the robot to complete the task in just one minute. This research aligns with ACE Robotics's "dual-brain" collaborative architecture and its real-world RSI technology roadmap, where the general model handles task understanding and skill generation, while the Kaiwu world model supports physical environment comprehension and strategy verification. Leveraging ACE-CaP-0 to enter the realm of code strategy development, ACE Robotics is exploring pathways for AI to autonomously learn and refine skills. As general models, embodied world models, and real-world RSI continue to merge, the company is accelerating the translation of understanding, planning, and generalization into fast, stable embodied skills. This approach shortens the timeline from autonomous exploration to practical execution, paving the way for robots to deploy in real-world industrial settings.