At the recent "RoboMind" Winter Camp for Intelligent Robot Developers Competition hosted by Huawei, the team from Beijing Institute of Technology (BIT) won the championship in the "clothes folding" track, earning the Excellence Award.

This competition witnessed nine top universities and research institutions competing against each other. The outstanding performance demonstrated BIT's strong capabilities in intelligent robotics and engineering practice.

This competition is the first national robot algorithm contest based on a domestic edge computing platform. The "clothes folding" contest, a representative challenge in the field of flexible object manipulation, required participating systems to autonomously complete tasks of flattening and neatly folding clothes under various conditions, such as smooth initial states, chaotic initial states, and changes in size and color. The focus was on evaluating the robot's ability to perceive flexible deformations during long-sequence operations, the continuity of action planning, and the overall robustness of the system.

The BIT team independently developed a lightweight unified decision-making model, which, due to its high degree of lightweight optimization, was well suited to the domestic edge computing platform. They completed model training using only about 20 hours of demonstration data, achieving excellent results.

The model integrates visual perception with the robot's motion state, allowing the robot to continuously perceive changes in clothing shape during operations and dynamically adjust subsequent actions.

Given the complexity, lengthy sequences, and high potential for error accumulation in the folding process, the team employed a segmented action, continuous inference method. This approach incorporates the results of executed actions into subsequent decision-making processes, achieving natural multi-stage transitions and effectively avoiding discontinuities between action segments. Even when faced with wrinkles, inversions, or variations in size and color, the system can still complete the entire process from flattening to folding.