Recently, a student team from the Shenzhen Loop Area Institute (SLAI), in collaboration with the Hong Kong University Multimedia Laboratory (HKU MMLAB), completed a rare domestic 24-hour continuous robot clothing-folding livestream. The entire process was led and narrated by the student team. The livestream was featured on the homepage recommendations of major domestic platforms, with cumulative views across all platforms exceeding 1 million, sparking widespread attention. This livestream not only demonstrated the robot's capability for long-duration, stable operation in the field of flexible object manipulation but also vividly illustrated the Institute's "3I" training characteristics, which emphasize international frontiers, interdisciplinary approaches, and industry-academia-research integration.

Students from SLAI led a 24-hour live stream on robotic clothes organizing, which was featured on the homepage of a major domestic streaming platform and has garnered over 1 million cumulative views.

A Round-the-Clock Robot "Endurance Race"

Unlike typical short promotional video demonstrations, the SLAI opted for a continuous 24-hour real-machine livestream. This approach fully presented the robot's execution of high-difficulty tasks such as folding and hanging clothes. The livestream did not merely showcase the technology in its entirety. It also exposed, reproduced, and resolved numerous issues encountered during actual operation. Ultimately, a positive and reproducible technical solution was publicly released.

In tasks full of uncertainty like clothing organization, the robot must perform precise operations while also possessing the intelligent adaptability to recover quickly from errors or even proactively retreat to advance. During the livestream, viewers could observe that when garments became extremely entangled, the robotic arm could strategically retreat, reset its steps, and turn a crisis into an opportunity. More notably, the audience could interact in real-time with the core R&D team, delving into technical details and offering suggestions for optimization.

Livestream Replay：

https://www.bilibili.com/video/BV1tcv4BgEid

Project Homepage：

https://mmlab.hk/research/kai0(Click "Read More" to visit.)

01

Why Has Clothing Organization Become a Global Arena for Top Institutions?

At the forefront of embodied intelligence, clothing organization, such as folding clothes, is becoming an invisible arena for the world's top labs. From Google DeepMind, Physical Intelligence, Stanford and Berkeley, research teams globally are exploring how to enable robots to understand and manipulate the highly uncertain physical world.

Folding clothes, as a non-deterministic task, involves input states—such as the shape, wrinkles, and position of garments—that are highly random. Fabric, a typical deformable object, introduces significant uncertainty. Therefore, folding flexible clothing is regarded as a core test scenario for evaluating a robot's comprehensive capabilities. It requires not only a high degree of coordination between visual perception and motion control but also the ability to transfer tasks and adapt strategies across different garment types. Furthermore, robots must possess multimodal generalization capabilities to handle open-world environments, which is essential for achieving true operational intelligence in the complex real world.

For a robot, folding clothes is nothing less than a comprehensive examination covering perception, reasoning, planning, and dexterous manipulation. In this livestream, the robot achieved a stable operational efficiency of organizing 60 garments per hour under typical conditions. This uninterrupted 24-hour run served as both a validation of the mechanical system's endurance and a comprehensive stress test of its perception, decision-making, and control systems.

A demonstration of the full process of organizing garments, the robotic arm continuously performs multi-step operations such as folding and hanging clothes. It maintains stable execution in scenarios involving flexible objects, demonstrating long-horizon decision-making and control capabilities.

During the livestream, scholars including Hao Zhao from Tsinghua University, Yao Mu from Shanghai Jiao Tong University, and Xin Jin from the Ningbo Oriental Institute of Technology were invited to participate in real-time discussions. They exchanged views on topics such as deformable object manipulation and the long-term reliability of embodied intelligent systems.

02

How to Build an Efficient and Reliable Clothes-Folding Robot?

A flowchart of the complete χ₀ model framework

To tackle the challenges of uncertainty, data distribution shift, and long-horizon task planning in flexible object manipulation, the team of Professor Ping Luo and Professor Hongyang Li from the Shenzhen Loop Area Institute has proposed a novel framework named χ₀. Its core innovation lies in rethinking the "value of data": through the art of alignment and fusion, it maximizes the utility of limited data. This enables industrial-grade performance, achieving a stable rate of folding 60 garments per hour in typical scenarios. The technical system of χ₀ is embodied in the following three key points:

Modal Consistency: Eliminating Errors at the Source

The data distribution during a robot's training often differs from the real-world operating environment, leading to "clumsy" manipulation. χ₀ addresses this through a modal consistency mechanism. By combining techniques like DAgger, spatiotemporal data augmentation, and inference-time control optimization, it aligns the training phase with actual operation. The result is that even when encountering unseen garment configurations, the robot can smoothly adjust its actions, significantly reducing the failure rate.

Model Arithmetic: A Shortcut in Weight Space

χ₀ introduces a model arithmetic method: models are trained on different data subsets separately, and their parameters are then merged through weight interpolation. Experiments found that the performance of this merged model can even surpass that of a model trained from scratch on the full dataset. This approach saves substantial computational resources, avoids the complexity of MoE architectures, and achieves a simple yet efficient boost in capability.

Phased Advantage Estimation: Endowing Robots with "Foresight"

Long-horizon tasks are prone to failure midway. χ₀ addresses this by decomposing complex folding tasks into key semantic stages. The robot predicts the advantage of the "current state" relative to the "target stage"—much like having a high-precision navigator. Instead of executing actions blindly, it can adjust its strategy at any time. This mechanism provides smooth, stable guidance signals, enabling the robot to complete long and complex operational chains.

03

How is the Industry-Academia-Research Model of Shenzhen Loop Area Institute Implemented in Practice?

Against the backdrop of top global labs competing to solve the clothes-folding challenge, the breakthrough by the Shenzhen Loop Area Institute is no accident. It is underpinned by the Institute's unique ecosystem for innovation and its deeply integrated industry-academia-research model. Conducting the research in Shenzhen provides technological support for the local robotics industry while also aligning with the national strategy to promote the entry of embodied intelligent robots into communities and homes.

In October 2025, five national departments, including the National Development and Reform Commission, jointly issued the Action Plan for Deepening Smart Community Development and Advancing City-Wide Digital Transformation. The plan proposes to "explore and promote the entry of embodied intelligent robots into communities and households," providing policy guidance for deploying technology in home scenarios. This transforms the "last-meter home scenario" into a national-level research topic, which resonates with the Shenzhen Loop Area Institute's research direction in domestic flexible object manipulation and continuous robot control.

Shenzhen Loop Area Institute Showcases Robotic Innovations for Home Scenes at 2025 University Sci-Tech Achievements Trade Fair in Guangzhou.

Strategic Advantage of Hetao's Cross-Border Integration Provides an Ideal Testing Ground for Industrial Challenges

The Shenzhen Loop Area Institute is located within the Hetao Cooperation Zone. This position enables it to integrate international academic resources with Shenzhen's industrial chain strength. To address challenges such as high model data costs and distribution shifts, the team proposed algorithmic innovations including "modal consistency" and "model arithmetic." These innovations reduce reliance on data and computing power, providing support for industrial reliability and engineering implementation. This R&D direction aligns closely with national policies promoting the deployment of robots in home scenarios, offering practical experience for achieving continuous progression from the laboratory to real-world home environments.

The "3I" Training Model Fosters Teams Capable of Solving Interdisciplinary Challenges

The Institute employs a "3I" training philosophy—International, Interdisciplinary, and Industrial. This approach breaks down disciplinary barriers and combines expertise from multiple fields such as computer vision, robotics, and control theory. It enables teams to propose innovative methods for complex, unstructured scenarios. For instance, "phased advantage estimation" grants robots long-term planning capabilities. Students take on core R&D and execution roles in high-difficulty technical tasks, thereby building a talent pipeline in advance for national home-application scenario research.

Focusing on Application and Ecosystem to Enhance Shenzhen’s Robotics Industry Competitiveness

The Institute has chosen the manipulation of flexible household objects as a breakthrough point. This closely aligns with Shenzhen’s service robotics and smart home industry layout. The framework provides an efficient and robust technical pathway, helping enterprises shift from reliance on sheer computational power toward intelligent precision. It accelerates the transition of technology from the laboratory to the industrial chain and into home environments. Through its integrated industry-academia-research model, the Shenzhen Loop Area Institute closely links its scientific research outcomes with policy direction. This aims to secure a strategic high ground in embodied intelligence and enhance the resilience and competitiveness of the industrial chain.

It is precisely within this innovation zone of the Greater Bay Area, which blends a global perspective, cross-border resources, and fertile industrial ground, and under the guidance of the "3I" training philosophy, that outcomes like the χ₀ framework—possessing both academic foresight and transformative industrial potential—can emerge. This is not merely a technological breakthrough. It represents a novel practice by Shenzhen’s new type of educational and research organization in translating national policy into cutting-edge research for home applications.