In recent years robotics firms backed by AI have banked billions of dollars as they attempt to bring the same model-driven breakthroughs seen in large language models (LLMs) into physical machines. The IPO of Unitree, one of China’s largest robot makers, seemed to validate that bet: the company was briefly valued at about $66 billion on China’s NASDAQ equivalent—but then saw its valuation cut in half after market realities set in. While the mechanical performance of these robots has improved, many still lack the intelligence to perform valuable, real-world tasks.
The Data Dilemma and Simulation Gap
At the core of the sector’s challenges is the data problem. High-quality, diverse training data is still hard to obtain for robotics—especially for tasks involving manipulation of objects. Developers are experimenting with richer datasets and novel training methods, along with reinforcement learning environments, to improve performance. An organizer of the Actuate conference—which now draws some 1,500 people, up from a fraction of that in 2023—warned that physical AI today resembles the GPT-2 era: lots of early promise, but still a long road to general intelligence. Part of the solution will be more compute and simulation, backed up by specialized GPUs that can produce high-fidelity environments.
Verticals vs Generality: A Strategic Trade-off
Some robotics firms are targeting niche verticals where performance and reliability can be driven by focusing on specific tasks. Industries like industrial automation, solar farm maintenance, construction, and excavation are already seeing real-world deployment. Take Bedrock, which automates excavators, and Agility operating in industrial environments—each working within clearly defined physical constraints rather than striving for undirected general-purpose humanoids. Generalist robots still largely exist in labs.
Other firms advocate co-designing hardware and AI from the start to navigate this tension. Executives argue that platforms should be flexible—hardware-agnostic where possible—to adapt to advancing sensors or novel components. At the same time, vertical focus provides usable data and recurring revenue, key inputs for training better embodied intelligence systems.
What’s Next: Moments That Matter
For many in the field, the awaited ‘‘ChatGPT moment’’ for physical AI would mean consumer-visible robot autonomy—machines that reliably do manipulation tasks out of the box, or vehicles that perform eyes-off driving affordably. One company license models to automakers to aim for eyes-off autonomy in a car for under $1,000 in hardware—a milestone that could signal a shift from laboratories to everyday life. The leadership gap may still be wide, but the few players that turn narrow-task deployments into scalable, general capabilities will move fastest.
Some doubt there will ever be an instant GPT-like leap in robotics. One entrepreneur suggested the real benchmark will be when consumers can buy a robot that delivers useful everyday functionality reliably—an Apple II or IBM PC moment for robotics rather than a ChatGPT flash of attention.
Where things stand: Robotics is entering a phase where physical AI firms must scale beyond mechanical novelty—they need data, simulation, and domain-specific use cases. Generality remains distant. Getting robotics to have meaningful consumer impact means building the infrastructure and reliability first.
Why This Shift Matters
Birth of modern robotics AI mirrors earlier waves of deep learning: vast promise, early demos, heated investment—but limited real-world impact until data, compute, and use cases aligned. The current moment is one of reckoning. Companies leaning into vertical applications risk being overtaken if general-purpose models pull ahead, but those pushing for generality before maturity risk building solutions with little market appeal. How firms balance breadth and depth today will determine who shapes the next decade of physical AI.
