At the 2026 World Robot Congress (WRC), held from August 19-23 at the Beijing Yichuang International Convention and Exhibition Center, a special session on "AI Large Models Empowering Robots and Embodied Intelligence Industry" took place alongside the main exhibition. Tao Shuai, Co-founder and CTO of Beta Infinite, delivered a keynote speech at this event. The following is the transcript of his presentation.
Tao Shuai began by introducing himself and expressing his gratitude for the opportunity to share his team's insights on embodied technology. He framed his presentation around the theme "Human-Centric: A New Paradigm for Embodied Technology in Human-Robot Coexistence Scenarios." He observed that the next wave of growth for AI and large models will inevitably extend beyond phones and screens, venturing into the physical world. He characterized "AI + Robotics" as potentially the most significant Beta growth opportunity of our era.
While 2026 is widely considered the inaugural year for embodied intelligence applications across various industries, Tao Shuai noted that the potential use cases span numerous sectors, from entertainment and performance to enhancing productivity in manufacturing. However, he clarified that Beta Infinite specifically focuses on consumer-grade applications. Their core mission is to serve people's lives rather than replace them, with the ultimate goal of creating genuinely useful robotic companions. He emphasized two key pillars: the robot must be useful, meaning it can perform tasks, and it must be a companion, meaning it can understand both the environment and the people within it.
Tao Shuai elaborated on the landscape of embodied intelligence, describing it not as a single ascending ladder but as a terrain of multiple, equally complex peaks. Each distinct scenario presents its own unique set of challenges, requiring a full-stack technical approach tailored to its underlying logic. For consumer-grade applications, this necessitates a native technological paradigm, rather than assuming that success in one domain (like industrial automation) automatically translates to another (like home assistance). He divided the overall embodied scenario into two main categories. The first is "task-centric," which operates in standardized environments with a limited number of SKUs. Here, the focus is on maximizing accuracy, cycle time, and reliability for single-point tasks. The second is "human-centric," which deals with the unscripted and unpredictable nature of human living spaces. To create long-term value in these open-ended environments, robots must be capable of executing long-horizon, complex tasks rather than simple, repetitive ones.
This fundamental difference in scenarios demands a complete restructuring of the technology stack, from the "brain" to the "cerebellum," operating system, and data infrastructure. For task-centric environments, the AI brain requires high certainty and can rely on a fixed set of instructions. In contrast, a human-centric, consumer-grade environment demands a robot brain with excellent memory of both the scene and the individuals, the ability to adapt to dynamic changes, strong generalization, and the capacity to interpret ambiguous human instructions. This extends to the hardware design itself, which must be safe, user-friendly, and cost-effective for consumers. From day one, Beta Infinite has positioned itself to build a native technology stack specifically for these consumer-grade, human-centric environments, built on three core principles: personalization, autonomous evolution, and human-robot friendliness. The first principle, personalization, requires the embodied brain to have robust multimodal memory (spatial, temporal, and person-specific) and a "world model" to understand and adapt to its environment. The second, autonomous evolution, addresses the immense data requirements for achieving ChatGPT-level intelligence in the physical world. Tao Shuai argued that pre-collected data or static data farms are insufficient; the only path forward is a self-closing data flywheel generated from deployed robots, enabling continuous learning and model improvement. The third principle, human-robot friendliness, dictates that the entire system, from the OS to the hardware, must be designed for safe interaction, proactive task planning, and consumer appeal. He stressed that a complete and satisfactory solution requires all three of these elements working in concert.
A cornerstone of Beta Infinite's technology stack is memory, which Tao Shuai identified as the essential capability that transforms a robot from a simple tool into an intelligent companion. He differentiated embodied memory from the memory used in typical AI agents. In the physical world, a robot perceives through multiple sensors (vision, hearing) and accumulates historical action experiences, creating a complex, dynamic, and spatio-temporally rich memory system. The key challenges are determining what to store, how to integrate this memory with the robot's "brain" (whether loosely or tightly coupled), and how to evaluate and refine the memory over time. To address this, Beta Infinite has developed a personalized reward system based on real-world physical feedback, using reinforcement learning to close the loop on memory storage, evolution, and brain integration. Testing in real environments has shown significant improvements in user experience, effectively giving the robot a "hippocampus." For example, in household tasks like finding objects (e.g., a cola or keys), the robot uses its memory to infer spatial relationships and plan its actions. It can even deduce likely locations for unseen items based on its past understanding of the environment. While minute- or hour-level memory is relatively easy, Tao Shuai highlighted the immense challenge of managing memory on a weekly, monthly, or yearly scale. Beta Infinite's current system is progressing from week-level to month-level memory, with the goal of achieving year-level memory by the end of the year.
On the topic of operational intelligence, Tao Shuai emphasized that generalization and long-horizon task execution are critical for consumer-grade robots. He predicted that future foundational models for embodied intelligence, such as VLA (Vision-Language-Action) and world models, will converge into a unified architecture, rather than remaining separate paradigms. This unified model would incorporate language, visual understanding, and prediction capabilities, similar to how the Transformer architecture became a standard. Beyond the model architecture, Beta Infinite is focusing on test-time learning and continuous evolution, enabled by a robust data flywheel. This combination allows the robot to adapt to novel objects and situations not present in its pre-training data—a necessary capability for entering diverse households across the world, where item appearances can vary dramatically. Tao Shuai cited the recent release of their v0.1 model, which successfully completed autonomous, long-horizon mobile manipulation tasks of over ten minutes in a real home environment. The model demonstrated fine manipulation skills, spatial understanding, and autonomous recovery from failures. For instance, when a book obscured a Lego block during a cleanup task, the robot used its memory and reasoning to infer the block's location, move the obstruction, and complete the primary objective. These capabilities were achieved end-to-end with a single model.
Finally, Tao Shuai argued that the future competitive landscape will not be defined by a single model or neural network, but by complex system architecture and integration capabilities. For embodied intelligence to truly integrate into human environments, it requires an open, agentic operating system—an "AgenticOS" for the physical world. Beta Infinite is building its framework from the ground up as an open-world physical agent, from hardware abstraction to high-level task planning. This system will first be refined through their own products, but Tao Shuai expressed a clear commitment to eventually open-sourcing it to foster a thriving ecosystem and collaborative development across the industry. He concluded his presentation by thanking the audience.
Comments