The race to build smarter robots just entered a new phase, and it's not about better hardware anymore,it's about artificial intelligence that can actually understand the physical world. Major AI companies like DeepSeek and Anthropic are now investing heavily in robotics startups, signaling that the next frontier of AI isn't in data centers but in factories, warehouses, and homes where robots need to think for themselves.

What's the Real Bottleneck in Physical AI?

For decades, roboticists have faced a frustrating paradox. Robots can play chess and solve complex math problems with ease, but teaching them to walk smoothly or pick up an object without dropping it remains extraordinarily difficult. This challenge, known as Moravec's paradox, reveals the core problem: hardware has been ready for over a decade, but the "brain" that lets robots understand and adapt to the physical world is still far behind.

The bottleneck isn't mechanical anymore. It's artificial intelligence. Specifically, it's the generalization problem, which refers to a model's ability to handle new situations it never encountered during training. Current embodied AI systems struggle with this, meaning a robot trained to pick up boxes in one warehouse might fail in a slightly different environment.

Unitree Robotics, the global leader in humanoid robot shipments with over 5,500 units delivered in 2025, openly acknowledged this weakness in its recent IPO filing. The company admitted it had historically focused on perfecting the robot's body and motion control while underinvesting in the "brain," or embodied large models that enable autonomous decision-making.

How Are AI Giants Entering the Robotics Game?

The strategic moves are unmistakable. DeepSeek, a leading Chinese AI firm, secured a 3-year lock-in period as a strategic investor in Unitree's recent IPO, signaling a commitment to long-term collaboration on three fronts: general artificial intelligence research, high-performance general robots, and large AI models. Meanwhile, Anthropic, another major AI player, reportedly explored acquiring Physical Intelligence, a robotics startup focused on embodied AI.

These aren't casual investments. They represent a fundamental shift in how the AI industry views the future. Jensen Huang, CEO of NVIDIA, has outlined AI's evolution in four stages: Perceptual AI, Generative AI, Agent AI, and Physical AI. He's declared that the "ChatGPT moment" for Physical AI has arrived, meaning the technology is reaching a tipping point where rapid commercialization becomes possible.

Unitree is raising 6.099 billion yuan (roughly $850 million USD) in its IPO, with plans to invest half of that directly into developing embodied large models, the AI systems that teach robots to think and learn autonomously. This represents a dramatic strategic pivot for a company that built its reputation on hardware excellence.

What Does This Mean for the Future of Robotics?

The convergence of AI expertise and robotics hardware is accelerating real-world applications. A global forum scheduled for September 7 in Seoul will bring together experts from industry, academia, and research institutions to discuss how Physical AI is reshaping manufacturing, logistics, agriculture, and healthcare. The event will feature live demonstrations of KAIROS, a Korean AI humanoid robot under development by the Korea Institute of Machinery and Materials (KIMM), designed for applications ranging from household assistance to disaster response.

The applications being showcased span multiple industries and use cases:

• Autonomous Delivery: Cooperative systems where drones and ground robots work together to transport goods across land, sea, and air
• Agricultural Monitoring: Autonomous robots that navigate open-field orchards to detect and monitor pests and diseases with precision
• Last-Mile Logistics: Delivery robots capable of seamlessly traveling inside and outside buildings by interfacing with shared entrances and elevators
• Healthcare Support: Smart robotic prosthetic hands and legs that assist users by responding to their intentions and movements
• Industrial Manufacturing: Robots performing complex assembly tasks autonomously, such as completing joint motor assembly without human intervention

Unitree has already begun testing its self-developed embodied large models in real-world scenarios. Its UnifoLM-X1-0 model has been piloted in its own factory, independently completing joint motor assembly tasks. In May, the company released WVLA2.0, which integrates world models with vision-language-action capabilities, enabling long-sequence autonomous planning and multi-step physical prediction. By July, UnifoLM-OminiA-0.3 was focused on home healthcare tasks, independently handling complex operations like object manipulation, sorting, and equipment control.

"Physical AI is a key technology that combines artificial intelligence with robotics and mechanical engineering to address real-world challenges not only in manufacturing, but also in agriculture, logistics, healthcare, and caregiving," stated Seog-Hyeon Ryu, President of KIMM.

Seog-Hyeon Ryu, President, Korea Institute of Machinery and Materials

How Are Companies Building Better Robot Brains?

The technical approaches are diverse, but the goal is unified: create AI models that let robots learn from experience and generalize to new situations. The industry is exploring three main technical routes: hierarchical large models, end-to-end large models, and world models. While these approaches differ, they all pursue higher generalization capabilities, which has become an industry consensus.

Unitree has adopted a two-pronged strategy combining self-research with strategic partnerships. On the self-research side, the company is developing both World Model-Action (WMA) and Vision-Language-Action (VLA) model routes. On the partnership side, Unitree is collaborating with NVIDIA and DeepSeek to accelerate development. At NVIDIA's GTC conference in June, Jensen Huang demonstrated the Unitree H2 Plus, a humanoid robot built on NVIDIA's Jetson Thor computing platform and powered by the Cosmos 3 model.

The competitive landscape is intensifying. Other robotics companies, such as Galaxy Universal, are purchasing Unitree robots to collect real-world data and train their own embodied large models, like the "Galaxy Star Brain." This creates a virtuous cycle where hardware leaders like Unitree become suppliers to AI-focused competitors, forcing them to accelerate their own AI development to maintain competitive advantage.

Why Does This Matter Right Now?

The timing is critical. Hardware capabilities have plateaued; further incremental improvements in motors, sensors, and mechanical design yield diminishing returns. The real breakthrough will come from AI systems that enable robots to perceive, reason, and act autonomously in unpredictable environments. Companies that crack this problem will dominate the next decade of robotics, just as companies that mastered large language models dominated the past two years of AI.

The strategic lock-in periods and acquisition attempts signal that major AI firms believe embodied AI is no longer speculative. It's becoming essential infrastructure. A robot that can learn from experience and adapt to new tasks is exponentially more valuable than one that simply executes pre-programmed motions. That shift from rigid automation to adaptive intelligence is what's driving the current wave of investment and partnership.