RoboParty releases source code for its humanoid robot RP1, including mechanical design and control software

A team from the Institute of Reliable Embodied Intelligence at Fudan University has introduced RoboSkill—a framework that enables general-purpose robots to retain acquired knowledge. Instead of collecting new demonstration data for every task, the agent first explores the task and then saves an effective method as a skill consisting of written instructions and executable code.

The research, titled 'Explore, Execute, Evolve: A Skill Acquisition and Reuse Loop for Embodied Agents,' was published on arXiv, and the corresponding code is available on GitHub.

The work represents a paradigm shift in robotics. Models that integrate vision, language, and actions traditionally require fresh data from each deployment site, and data collection from robots is costly due to the need for teleoperation, scene resets, and demonstration repetition.

The team found that modern multimodal agents are capable of observing, reasoning, and writing control code directly during operation. In the LIBERO-10 simulation environment, with a four-hour budget and multiple attempts, the best agents achieved one hundred percent success. When operating on a real robot, the strongest agent successfully completed 91.7 percent of simple tasks and 80 percent of complex ones.

The main challenge remains speed. Since agents cyclically go through observation, reasoning, action, and verification stages, the reasoning process between short arm movements can take minutes. The average duration of simulation runs was 94.5 minutes, and successful task completion on a real robot required 20 to 27 minutes.

RoboSkill addresses this issue in two ways. First, it incorporates tactile feedback. Pressure, force, and torque readings inform the agent about gripper contact, eliminating the need for constant image checking. After adding tactile data, the average simulation time for one agent decreased from 25.35 minutes to 15.93 minutes, and the pipette extraction task on a real robot was reduced from 33.63 minutes to 20 minutes.

Second, the framework accumulates skills. After completing a task, the agent analyzes what happened and documents the effective method, subsequently updating it based on subsequent results. For example, the method of determining table height via contact can be used in a new scene, even if the height needs to be measured again. Skill reuse allowed the average simulation time for one agent to decrease from 94.5 minutes to 35.9 minutes.

The most significant contribution was saving code alongside text. The transition from text-only skills to those incorporating both text and code reduced the average time for all four tested agents. For one of them, the time decreased from 67.1 to 39.9 minutes, and the first-attempt success rate increased from 15 percent to 67.5 percent. Overall, the team reports an increase in first-round success on LIBERO-10 from 12.5 to 25 points, a reduction in simulation time by 7.6 to 72.4 percent, and an increase in real-robot success by 8.3 points. The researchers note that skills can also be transferred between different tasks and different agent models.