At the "Physical AI and Embodied Intelligence Robot Innovation Ecosystem Exchange", a special session of the China International Fair for Trade in Services held in Beijing on September 11, 2026, Fan Jianping, a distinguished professor at Fudan University's Trustworthy Embodied Intelligence Research Institute and founder of OpenFriday, shared his insights on the field's latest developments.

Fan began by emphasizing that future robots will not be oversized remote-controlled toys; they must possess an intelligent soul, or agent, at their core. As robots will need to operate not only in cyberspace but also execute actions in the physical world, this fundamentally makes them cyber-physical agents.

To illustrate the dual nature of these systems, Fan introduced the concept of a "dual model for embodied cognition." This architecture requires two distinct "brains": one for handling traditional agent functions such as memory, tool usage, data access, and long-horizon planning, and another dedicated to controlling the robot's physical movements and actions in real-time.

Over the past few years, his team has developed a comprehensive development platform for cyber-physical agents. This platform encompasses crucial components including agent development tools, model fine-tuning processes, data and knowledge management, multi-agent collaboration protocols, and inference acceleration technologies.

This platform is designed to support both "agent-native devices" like robots, AI-enabled phones (AI Phones), and AI personal computers (AIPCs), as well as smart solutions such as smart manufacturing. The core belief is that future robots must move beyond being simple toys and integrate into the workforce within factories and workshops to provide genuine value.

Discussing the evolution of hardware, Fan drew a parallel between technological advancement and intelligence development. Just as internet PCs followed the PC era and smartphones emerged after mobile and wireless communication advances, the current AI boom is set to completely reshape hardware into what he terms "AI-native devices." Within this new paradigm, the hottest trend is the agent, paving the way for "agent-native devices" where intelligence is fundamentally supported by an embedded agent.

What defines an agent-native device? It requires a more powerful "brain" in the form of advanced agent software and more robust computational hardware to run it. Crucially, these devices must be developed through hardware-software co-design, with three layers (agent, OS, and hardware) evolving in tandem. The agent layer is responsible for reasoning, decomposing user intent, and performing tasks in cyberspace. The OS layer handles model inference acceleration, task scheduling, and memory management, while also potentially incorporating advancements like an "Agent Harness." The bottom layer, the chipset, provides the necessary raw computing power.

This convergence is possible today because both software (agent intelligence and knowledge density) and hardware (computing capabilities) are improving in tandem, making it feasible to develop powerful yet resource-constrained models that can run on local devices with sufficient capability. The primary forms of these agent-native devices include PCs, phones, and crucially, robots. These robots, powered by an embedded agent, will be able to understand human needs, execute corresponding physical actions, and adapt to diverse environments and tasks over time.

Shifting focus to applications, Fan explained how his platform supports smart manufacturing, given China's strong manufacturing base. The entire smart manufacturing stack spans engine control for security-sensitive companies, demand forecasting, production planning, quality control, and even complex logistics like shifting semi-finished products to Mexico to circumvent tariffs, all while optimizing costs.

Fan focused on two key areas: demand forecasting and quality control. In demand forecasting, supply chain experts generate initial hypotheses about how factors like weather could influence orders. An intelligent agent then creates a prediction based on these hypotheses. Experts can review this prediction against historical experience, adjust hypotheses, and have the agent re-evaluate, iterating until a reasonable forecast is achieved for guiding production. He noted that this demonstrates that AI won't replace humans, but will work alongside them, with people always making the final decisions.

For quality control, he highlighted defect detection as a critical issue. His team has developed a defect detection agent that can be commanded by workers on the production line, significantly improving quality assurance. The ultimate goal is to establish an AI-native organization where AI employees, both digital employees operating in cyberspace and physical robots on the production line, collaborate seamlessly with human employees to solve pressing manufacturing challenges.

Despite the promising future, Fan outlined significant challenges. Currently, many physical devices remain remote-controlled toys, capable of solving only specific, well-defined problems. A primary hurdle is data. Unlike the internet, which amassed vast data before the large model boom, embodied intelligence data is extremely difficult and expensive to collect, forcing many companies into costly data acquisition.

Furthermore, cyber-physical agents must plan for complex long-horizon tasks spanning both cyberspace actions and physical actions. Coupled with the highly variable physical environment, this leads to poor generalization capabilities for AI systems operating in the real world. Another major barrier is the stringent constraint of power consumption and on-device computing power for edge devices, which makes implementation incredibly challenging.

Finally, as we contemplate AI and human employees working together, ethical considerations become paramount. This necessity is why his research institute is named "Trustworthy Embodied Intelligence," underscoring the commitment to not only advancing the technology but also addressing the profound ethical questions it raises. He concluded his presentation by thanking the audience.