On April 28, at the “AI Riding the Tide—New Trends of 2025 Jiazi Gravity X Technology Industry” conference, Zhao Yue, founder and CEO of SEER Robotics, shared the company’s latest developments in the field of embodied AI and elaborated on the future industrial landscape of intelligent robots under the embodied AI revolution.

Embodied AI is leading technological breakthroughs and reshaping the application boundaries of intelligent robots. Where is the business implementation of embodied AI as it transitions from technological imagination to industrial reality? As the world’s largest intelligent robot company centered on control systems, SEER Robotics has provided an answer to the implementation of embodied AI with a “new-old integration” mindset.

Five Development Stages of Intelligent Robots

The first stage is “Software Programming and Hardware Scene-Specific.” Fixed programming is required for each scenario, and hardware is designed entirely around specific needs. Typical applications include industrial robotic arms and automated guided vehicles, which can only operate in preset fixed scenarios. The industry characteristic at this time is “one machine for one scene, one action for one programming,” with high development costs and a lack of flexibility.

The second stage evolves to “Software Scene-Specific and Hardware Scene-Specific.” With the increasing demand for flexibility and scene autonomy, collaborative arms and autonomous mobile robots have begun to emerge. Software needs to be fine-tuned according to different scenarios, and hardware is customized based on specific needs. For example, forklifts require specialized pallet fork mechanisms, and material box robots are equipped with clamping devices.

We are now entering the third stage of “Software Generalization and Hardware Scene-Specific.” Driven by AGI technology, software is moving towards generalization, but hardware still needs scene-specific customization. For example, the robot juggling handkerchiefs performance shown in the Spring Festival Gala appears to exhibit high flexibility, but in fact, the action stability is achieved through fishing line traction. In industrial scenarios, considering cost, efficiency, and safety, hardware customization based on scenarios will continue to exist for a long time.

Even if we enter the fourth stage in the future to achieve “Software Generalization and Hardware Generalization,” with general humanoid robots becoming mainstream, they will still coexist with specialized equipment for a long time.

In the fifth stage, we envision a “return to carbon-based,” where robots may possess the adaptive ability like liquid metal, achieving morphological evolution that adapts to the environment and moving towards the deep integration of carbon-based life and silicon-based intelligence.

Three Challenges in the Implementation of Embodied AI

Challenge One: Data Dilemma

Humanoid robot remote operation data collection is difficult and the quality is questionable. For example, the lack of wrist joints in humanoid robots leads to an essential difference between the collected actions and natural human behavior, resulting in data that is detached from real scenarios.

Challenge Two: Feedback Limitations

Human skin has a large number of sensory units that achieve precise tactile communication. However, humanoid robots only have basic feedback and lack various types of tactile, force, and temperature sensing, resulting in relatively weak feedback.

Challenge Three: Low Controllability

Humanoid robots have many joints and the difficulty of multi-joint coordinated control increases. A single action often requires screening a large amount of training data to ensure reliability.

Two Innovative Paths of SEER Robotics

Path One: New Technology + Old Product = Embodied AI Forklift

We have applied technologies such as multi-layer semantic maps, end-to-end navigation, and VLA/reinforcement learning to the automatic forklift scenario to create an embodied AI forklift.

● Constructing multi-layer semantic maps, the underlying point cloud ensures the reliability of environmental perception, and the middle layer of environmental semantics endows the system with scene understanding capabilities to enhance generalization.

● The end-to-end navigation technology draws on the experience of autonomous driving to achieve real-time navigation of forklifts in any indoor environment without predefined paths.

● VLA enhances generalization capabilities, and reinforcement learning improves reliability. We leverage existing massive manual operation data and the low degree of freedom characteristics to verify the feasibility of the technology.

Path Two: Old Technology + New Product = Embodied AI Controller

● The SRC-5000, launched by SEER Robotics in 2024, is the world’s first integrated embodied AI controller. It is provided with a complete set of industrial chain tools to help customers quickly develop various embodied intelligent robots, such as humanoid robots and robot dogs. Now, the SRC-5000 has been implemented at customer sites.

● The SRC-5000 integrates dedicated AI chips and co-processors in a heterogeneous manner, achieving high stability, high real-time, and high performance. It has industry-leading integration and has realized Whole-body Control first, breaking through the technical bottleneck of hand-eye-foot coordinated control and building a complete “brain + cerebellum” perception and decision-making system.

● Based on the SRC-5000, we have successfully applied multi-modal models such as visual semantic maps and end-to-end to both intelligent forklifts and wheeled humanoid robots, achieving a certain degree of improvement in generalization capabilities.

Build Your Own Robot Fleet Within Days!

SEER Robotics always practices the mission of “Build your own robot fleet within days！” eliminating industry barriers from three dimensions: development, acquisition, and use.

● Making robot development barrier-free: We provide industry-leading robot control systems that have been integrated with more than 300 core component brands on the market, truly achieving plug-and-play for components and helping enterprises easily develop their own intelligent robots.

● Making robot acquisition barrier-free: We have built a one-stop robot selection platform—Nebula Platform, which gathers more than 1,000 different types of robots developed by partners. Customers can customize and select robots on the platform in a one-stop and online manner, quickly obtaining the robots they need.

● Making robot use barrier-free: We offer All-in-One software tools that seamlessly connect with robot systems and customer upper-level systems, and provide low-code development modules. Engineers can easily drag and drop to meet customized needs.

To promote “Build your own robot fleet within days！” SEER Robotics has established a dual-flywheel drive model of “technology flywheel + platform flywheel.”

● Platform Flywheel: SEER Robotics focuses on the research and development of robot brain control systems, aiming to lower development barriers and enable more enterprises to develop their own robots. This creates a diverse and open robot ecosystem, providing downstream integrators and end customers with more robot options and driving platform development.

● Technology Flywheel: Through a variety of scene-specific embodied hardware, we continuously meet the needs of end customers, achieving rich data and scene accumulation. Combined with AGI technology, we enhance controller performance, truly creating highly reliable and generalizable universal software and accelerating the evolution of SEER Robotics’s AGI brain.