Xu Huazhe of PoKe Robot: Vision for Robots Serving the Public May Be Realized Within 3-5 Years

　　Shanghai Securities News reported to China Securities News (Reporter Li Yanzheng): How far is general-purpose robotics from the lives of ordinary people? Beyond various flashy demonstrations, many robots are still trapped in specific scenarios, able only to repeat a few preset actions. Once faced with unfamiliar objects or variable work environments, the task success rate drops significantly.

　　Regarding this, Xu Huazhe, Assistant Professor at the Institute for Interdisciplinary Information at Tsinghua University and founder of PoKe Robotics, stated in an exclusive interview with a Shanghai Securities News reporter that the industry cannot be satisfied with robots repeating a single action 10000 times; the core of true general-purpose robots is completing ten thousand completely different actions. With the rapid iteration of embodied large models, zero-shot generalization it is moving from theory to practice; within 3-5 years, the vision of robots serving the public is expected to gradually be realized. However, the industry still needs to leap beyond the closed loop of demonstration demos and overcome the challenges of real-world generalization across different bodies, scenarios, and tasks.

　　The core of general-purpose robots is multi-task generalization

　　In industrial scenarios, traditional robotic arms are good at repeating the same action tens of thousands of times to complete high-tempo repetitive tasks such as sorting and moving. In Xu Huazhe's view, while this type of capability possesses practical commercial value, there is still a significant gap from the goal of general-purpose robots. “Doing one action 10,000 times solves repetitive labor; doing 10,000 different actions is what”Embodied Intelligence The directions pursued, the underlying logic of the two is completely different.

　　Currently, machine learning is widely introduced in robot R&D, but many technical advances have not truly achieved generalization. He cited the case of overseas humanoid robots folding blankets as an example; the demonstration video effect is stunning, but in zero-shot real-world tests in ordinary homes, the overall success rate is not high. Robots perform excellently in fully debugged demonstration environments, but once they leave the training dataset and face changes in object shapes or home layouts, they easily fail, which is already a common pain point in the industry.

　　“Two years ago, the industry was still debating the pros and cons of traditional control versus learning paths, and end-to-end versus layered architectures; today, the debate has concluded, but generalization remains the mountain standing before the industry.” Xu Huazhe told reporters. The Pokexu Robotics team once completed a technical experiment of a robot making Mapo Tofu, achieving multiple difficult robot practical tasks such as belt tying. Currently, a certain degree of generalization has been achieved; what needs to be solved in the future is cross-embodiment generalization, cross-scene generalization, and cross-task generalization.

　　In actual industrial R&D, R&D personnel need to repeatedly collect scene data and iterate models. Even with continuous expansion of the dataset, it remains difficult to meet actual usage requirements. Project cycles can be as short as a few days, while large projects can reach several months, and a large amount of invalid waste data is generated during the process; meanwhile, engineering problems such as robotic arm overheating are encountered. “Relying on repeated data collection and model tuning makes it impossible to scale into households. Our ideal is that as long as the hardware itself meets the standard, once the user issues a natural language command, the robot can complete the task with zero-shot, no longer wasting a large amount of manpower to collect data.” said Xu Huazhe.

　　RoboAPI Will build the ChatGPT of the embodied world

　　Currently, GPT-Artra Demonstrates remarkable zero-shot creation capabilities, drawing paintings without any samples and even performing self-reflection and iteration. However, Xu Huazhe also pointed out its shortcomings: slow running speed, high costs, and relatively poor performance in dexterous operations in the physical world. Painting belongs to the arrangement of colors in two-dimensional space, while robot operation requires dealing with three-dimensional physical interactions, where variables such as contact force and object deformation are full of uncertainty; the difficulties of the two cannot be compared.

　　“Large models prove that zero-shot is feasible, but the embodied field needs its own foundation models for dexterous manipulation.” Xu Huazhe judged that embodied intelligence should refer to the large model API path to create the ChatGPT of the physical world. To this end, Pokex Robot launched the RoboAPI cloud brain interface, aiming to achieve “command and obey.” After developers connect the hardware, they do not need to collect or train models themselves; by relying on natural language instructions, they can drive robotic arms to complete various tasks such as sorting strawberries, folding clothes, pouring coffee, wiping tables, and opening cabinets.

　　Benchmarking horizontally against GPT-Artra,π0.5and other mainstream models, RoboAPI performs ahead in test tasks such as object transfer, pose and shape alignment, and environmental interaction, with a particularly prominent speed advantage, accelerating by approximately 20 times compared to GPT-Artra. Currently, RoboAPI has been opened to the public, and laboratories from multiple universities including Tsinghua Deep Institute, Boston University, Northeastern University, and NTU have completed real-world tests, reproducing zero-shot generalization capabilities.

　　This interface possesses strong cross-embodiment adaptation capabilities; in addition to its own hardware, it is also compatible with mainstream commercial robotic arms such as UR, and many developers have already conducted tests in laboratories. Xu Huazhe frankly stated that performance will experience some degradation when running across different hardware, which is also a direction for the team's continuous optimization. At the same time, he reminded developers that they need to avoid camera occlusion during use, and instruction descriptions should be as clear as possible; the system has built-in safety mechanisms and will directly refuse to execute instructions that harm humans.

　　The dawn of large-scale deployment of general-purpose robots has appeared

　　Behind RoboAPI is the Poke underlying model system as support. Xu Huazhe introduced that, in addition to basic language sub-task annotation, the team innovatively adopted the CEA annotation system, performing refined labeling for object contact points, contact time, and contact expectations, completing a cumulative amount of data labeling on the order of several billion times. Refined physical interaction labeling effectively reduces the randomness of robot actions and improves the certainty of execution in physical environments.

　　Looking back at industry development, Xu Huazhe stated that in the past, the market often questioned whether general-purpose robots could be implemented, but today this technical path should no longer be denied. Of course, technical issues regarding probability still exist and 100% perfection cannot be achieved, but zero-shot capability has already changed from a concept to a verifiable reality. He predicts that within 3-5 years, the industrial vision of robots serving humans is expected to gradually be implemented. In the future, robots undertaking venue services may become a landmark scenario for the maturation of general technology.

　　He called on more developers to join the general-purpose robot track; software practitioners do not need to undertake heavy data collection and model training from scratch, and can focus on upper-layer business development by relying on RoboAPI.

　　“General-purpose robots, they must be realized.” Xu Huazhe believed, AGI Not limited to text and images, only when intelligence truly possesses a physical body and understands the complex physical world can it release huge social value and reshape the labor modes of homes and factories.