Physical AI has become a major focus for venture investors, as companies try to bring the techniques behind Large Language Models into robotics. The promise is clear: machines that can understand the physical world well enough to perform valuable tasks, not just move impressively in a demo.
But the gap between robotics excitement and robotics utility remains large. At the center of that gap is data: how to collect it, manage it, simulate it, and turn it into models that can work reliably outside controlled environments.
Why investor enthusiasm is colliding with practical limits
The market’s appetite for robotics is visible in the scale of funding and public-market attention. Unitree, China’s leading robot maker, reached a $66 billion valuation after listing on China’s equivalent of the NASDAQ. This week, however, the company lost nearly half of its value.
Analysts pointed to a basic issue. Robot bodies are getting better, but many robots still do not have the practical intelligence needed to create value. In other words, movement is not the same thing as useful work.
That tension was clear at last week’s Actuate conference, a gathering for developers building AI brains for robots. According to organizer Foxglove, the event has tripled in size since it began in 2023 and had 1500 attendees. Foxglove helps physical AI model builders manage and visualize their data.
The excitement at the conference came with a warning sign. Avala, another physical AI infrastructure company, used its booth to promise a solution to “the robotics data crisis.”
The data problem holding robots back
The crisis is a shortage of high-quality training data for AI models. Building a robot that can perform any task remains far away, and end-to-end learning for specific jobs has not yet produced reliable commercial products.
Developers are now trying to follow the path of frontier AI labs. That means searching for more diverse data sets, testing different training approaches, and improving reinforcement learning setups. The work is less about a single breakthrough and more about building the systems that let robots learn from richer examples.
Harry Mellsop, a founder of Antioch, described physical AI as being in its “GPT 2 era.” Antioch builds simulation tools for model builders. His point was that the field may still need more data and compute before it moves beyond today’s limits.
One specific need is GPUs optimized for ray tracing, which are used to create high-fidelity simulations. Simulation matters because robots must be tested against complex physical conditions before they can be trusted in the real world.
Why autonomous vehicles are ahead
Autonomous vehicles are further along than many other areas of physical AI. One reason is that companies can collect relevant data from cars driven by people. Another is that driving mainly requires avoiding contact, while many humanoid and manipulation tasks require robots to interact with the physical environment directly.
Much of the current tooling for robotics model-building comes from autonomous vehicle companies. Foxglove, for example, was founded by former employees at Cruise, General Motor’s erstwhile self-driving effort.
Some companies that began with vehicles now see humanoid robotics as a logical research direction. Tesla is working on its Optimus robot. Wayve and Uber have also launched robotics labs focused on humanoid form factors as R&D efforts.
Alex Kendall, the CEO of Wayve, told TechCrunch that the field should begin with vehicles. He said manipulation robotics is like self-driving five years ago, while data infrastructure, simulation, and ML ops infrastructure will likely be shared across different robotic systems.
Kendall also argued that it is too early to commit to a single hardware platform. Sensors and other components are advancing quickly, and a general model should not be too tightly tied to one embodiment.
The debate over general robots versus focused work
Not everyone agrees that a brain-first strategy is right for this stage. Théophile Gervet, the CEO of Genesis AI, said the field is too early for that approach. Genesis AI is a vertically-integrated humanoid robotics company that raised a $105 million seed round this year.
Gervet argued that there are still many opportunities to design hardware and AI together. His view connects to a larger question for physical AI companies: should they pursue general-purpose robots, or focus on specific jobs where deployment is possible now?
The source shows a clear split in the market:
- Gritt is building solar farms.
- Agility is deploying robots in industrial settings.
- Bedrock is operating excavators autonomously.
- General-purpose humanoids are still not getting out of the labs.
Gervet framed the customer problem directly: “No customer cares about the general purpose robot that works at 80% success rate,” he said. At the same time, he warned that building for a narrow vertical on top of weaker model capabilities could leave a company exposed when stronger general systems arrive.
Focused deployment has one major advantage. It can produce revenue while also generating real-world data. Bedrock CTO Kevin Peterson said the company is starting with excavation to understand the challenges of “manipulation in the wild,” while planning an intelligence layer that can extend across multiple construction machines.
What a breakthrough might really look like
Managing robotics data is itself a technical challenge because visual and lidar data are dense. Foxglove announced a new product this week built on Nvidia’s Cosmos open weight world model. The product lets engineers search that data with natural language queries so they can build evaluations and simulations faster.
The broader question is what physical AI’s ChatGPT-like moment would be. Kendall pointed out that the largest robot deployments in the world are still consumer vacuum bots. For him, the next major milestone would need to excite consumers, not just investors.
He offered one example: eyes-off autonomy for less than $1000 worth of hardware in a car. Wayve is licensing models to car makers in pursuit of that goal, which Kendall sees as a multi-billion dollar opportunity and a path toward a general embodied AI model.
Gervet’s version of the breakthrough is manipulation that works out of the box: a robot that can take natural-language instructions and perform basic physical tasks such as pushing, pulling, closing a laptop, or cleaning up a table with 80% plus reliability.
Foxglove CEO Adrian Macneil was more cautious about the comparison. “There will not be a ChatGPT moment for robotics,” he told TechCrunch. His reasoning was distribution: software can reach users quickly, while distributing machines in the real world is much harder.
Macneil instead pointed to a different kind of milestone: the moment when people can buy a home robot that starts doing useful and fun things. For physical AI, the breakthrough may not arrive as a sudden consumer software event. It may arrive when data, simulation, hardware, and deployment finally make robots useful enough to leave the lab.