Ahead of the highly anticipated World Artificial Intelligence Conference (WAIC) 2025, set to take place from July 26 to 28 in Shanghai, PsiBot, a leading player in China's embodied AI sector, is gearing up to make a striking debut. Among its showcase of cutting-edge technologies, two standout innovations – a mahjong-playing robot and a food-delivery robot – are poised to steal the spotlight, offering a vivid glimpse into how embodied AI is transitioning from lab breakthroughs to real-world applications.

A showcase of end-to-end embodied AI capabilities

PsiBot, known for its expertise in general embodied intelligence, VLA large models, and dexterous manipulation algorithms, will present a comprehensive suite of technologies at WAIC 2025. Beyond the headline-grabbing mahjong and delivery robots, the exhibition will feature core innovations such as the hierarchical end-to-end VLA model Psi R1, PsiBot H1 high-degree-of-freedom dexterous hand, PsiBot full-body robot and teleoperation systems. These technologies span retail logistics, industrial applications and entertainment, demonstrating Lingchu's full-stack capabilities from data and algorithms to hardware and scenario-based deployment.

"These two robots are a concentrated display of multi-scenario embodied intelligence," said Wang Qibin, founder and CEO of PsiBot. "The mahjong-playing robot can complete continuous mahjong games with on-site audiences, performing complex operations such as flipping, drawing, and sorting tiles with millimeter-level precision. It demonstrates strategic thinking in real-time battles with human players, dynamically builds decision chains, and independently makes game decisions like 'peng' (matching a pair) and 'gang' (four of a kind). Audiences can not only watch on site but also play against the robot in person, experiencing how it integrates visual, language, and action information to form a complete Chain of Action and Thought (CoAT)."

He added, "The intelligent food-delivery robot demonstrates a solution to the 'last 100 meters' delivery challenge. Facing diverse and easily deformable delivery objects like express bags and paper bags, the robot accurately identifies different shapes and placement postures. Its dexterous hand, as flexible as a human hand, can nimbly pass through handles and grip stably. With strong generalization and dynamic performance, it can handle deformed objects, safely and stably lifting bags of any size and weight. Equipped with a 7-degree-of-freedom robotic arm and a 6-degree-of-freedom dexterous hand, it can handle delivery tasks with a 3-kilograms load, providing a complete intelligent solution for food delivery and logistics scenarios."

Breaking down technical barriers: from mahjong tables to city streets

The mahjong-playing robot, a crowd-pleaser designed to highlight advanced AI capabilities, executes 30-minute+ games with millimeter-precision actions: flipping tiles, sorting and making strategic decisions like "peng" (matching a pair) or "gang" (four of a kind). Powered by Lingchu's proprietary CoAT framework, it integrates vision, language and motor skills to form a closed loop of reasoning, planning, and execution – true "thinking hand."

Chen Yuanpei, co-founder and head of reinforcement learning at PsiBot, noted that the mahjong-playing robot with autonomous reasoning ability is one of the cutting-edge technologies making its debut. "It realizes intelligent decision-making in complex strategy games, showcasing our breakthroughs in reinforcement learning that enable long-term complex task handling in open environments," he said.

Meanwhile, the food-delivery robot addresses the "last 100 meters" of logistics, a long-standing industry pain point. It navigates dynamic urban environments, handles flexible objects like crumpled paper bags and adapts to unexpected obstacles. 

"Our display of these two scenarios is deliberate. The mahjong game demonstrates the robot's ability to handle long-term complex tasks in an open environment, which requires the combination of strategic thinking, dynamic decision-making and precise operation – reflecting our technical breakthroughs in the field of L3 long-range dexterous operation. Food delivery, on the other hand, solves actual commercial pain points, showing the robot's ability to handle flexible objects, navigate complex environments and perform generalized grasping," Wang Qibin explained.

21-DOF dexterous hand with fine force control, high precision and high dexterity

The self-developed high-degree-of-freedom dexterous hand of PsiBot supports high-precision motion control and tactile feedback. It is capable of fine force control and stable execution for various operations such as single-finger, full-hand and multi-finger combination movements. It not only achieves human-hand-level trajectory accuracy and operational freedom but also features highly human-like motion continuity and interactivity, meeting the dual requirements of industrial deployment for fine operations and dynamic adaptation. 

Chen Yuanpei added, "We are also unveiling other first-time public technologies, including the dual-arm and dual-hand teleoperation exoskeleton glove that demonstrates precise human-robot collaborative control, and the high-degree-of-freedom dexterous hand technology that breaks through the limitations of traditional robotic arms. These technologies start from core challenges of embodied intelligence such as long-range planning and dexterous operation, echoing WAIC's focus on global AI frontier trends."

The Chinese path to embodied AI: differentiation through self-reliance

PsiBot's innovations underscore a distinct Chinese path in embodied AI – characterized by end-to-end in-house R&D and scenario-specific customization. Unlike fragmented global approaches, Chinese AI companies like PsiBot control every link in the value chain: from algorithms (Psi R1 model) to hardware (21-DOF dexterous hands) and real-world deployment.

"Our 'Chinese path' is reflected in several key aspects. Firstly, in technological path innovation: while global peers are still lingering in the L1 generalized grasping stage, we have broken through the industry watershed of L2 dexterous operation and established a leading position in the L3 long-range dexterous operation field. Our Psi R1 model has achieved 30-minute continuous mahjong games for the first time. This leap from simple action execution to real cognitive decision-making and long-range operation demonstrates the forward-thinking mindset of Chinese enterprises in AI technological paths," Wang Qibin emphasized.

"More importantly, our technological achievements stem from real customer needs and on-site task verification. From mahjong games to intelligent delivery, each scenario solves actual commercial problems. This application-oriented technological innovation path is the differentiated advantage of Chinese AI enterprises in global competition," he added.

World Artificial Intelligence Conference 2025 (WAIC): a global stage for AI competition

As WAIC 2025 prepares to host over 800 enterprises from more than 30 countries (over 50 percent international), AI companies like PsiBot symbolize China's rising role in AI leadership. The conference, which runs from July 26–28, 2025, spanning over 70,000 square meters across Shanghai's Pudong and Xuhui districts, focuses on AI infrastructure, industrial empowerment and global governance – areas where embodied AI is set to disrupt industries. 

Wang Qibin said that it is an inevitable trend to release more embodied AI applications for public life scenarios. "The value of embodied intelligence lies in its ability to truly integrate into people's daily life and work scenarios – from household services to commercial applications, from entertainment interactions to manufacturing. We will continue to use platforms like WAIC to showcase how embodied AI technology solves more practical problems, making technology truly serve people," he noted.

Chen Yuanpei believes that reinforcement learning will play a core role in promoting AI from the virtual to the physical world. "With its unique value-driven mechanism and trial-and-error learning ability, reinforcement learning plays a key role in the implementation of embodied intelligence. Just as large language models have achieved qualitative leaps through large-scale data training, reinforcement learning technology will gradually become the mainstream technical paradigm in the field of embodied intelligence through continuous iteration and optimization in actual scenarios," he said.