Physical AI is one of the hottest areas for venture funding, with companies raising billions of dollars to apply tools that provide large-scale language models to robotics.
That excitement helped lead to a major IPO for Unitree, a major Chinese robot maker, that valued the company at $66 billion after listing on China’s Nasdaq equivalent. But the bottom fell out this week, with the company losing nearly half its value. Analysts point to one obvious problem. That is, although robots’ physical capabilities have improved, they still lack the know-how to perform tasks that actually create value.
The excitement was palpable at last week’s Actuate conference, a gathering of developers building AI brains for robots. The event has tripled in size since it began in 2023, with 1,500 participants, according to organizer Foxglove. Foxglove is a company that helps physical AI model builders manage and visualize data.
The risks were also clear. Avala’s booth sign, another physical AI infrastructure player, promised to solve the “robot data crisis.”
The crisis is the lack of high-quality training data for AI models. Attempts to build general-purpose robots that can perform any task are still a long way off, and even using end-to-end learning for specific tasks has not yet delivered a product with reliable commercial performance. For developers, the answer is to better mimic the advances of cutting-edge AI labs. That means finding or creating more diverse datasets, trying different training plans, and finding better reinforcement learning scenarios.
Harry Melsop, founder of Antioch, a startup that develops simulation tools for model builders, suggests that physics AI is in the “GPT-2 era,” or pre-ChatGPT OpenAI models. Overcoming this problem will require more data and compute, especially GPUs optimized for ray tracing, which is used to create high-fidelity simulations.
Self-driving cars are the most advanced. This is partly because relevant data can be collected from cars driven by humans, and partly because the main task is to avoid contact rather than to manipulate the physical environment. Many of the tools for building models are provided by self-driving car companies. Foxglove, for example, was founded by former employees of General Motors’ former Cruise company, which was working on self-driving cars.
And now these car companies are increasing their bets that investments in ML tools will help them compete with dedicated humanoid manufacturers. Tesla has already tried this with its Optimus robot, and now both AV-focused Wave and ride-sharing giant Uber have launched robotics labs focused on humanoid form factors as research and development efforts.
“I think we need to start with vehicles…operational robotics is what self-driving was five years ago,” Wayve CEO Alex Kendall told TechCrunch. “The data infrastructure, simulation, and ML operational infrastructure will probably be shared, but the specific world model of the simulator will be different after training. There will be much more commonality than nothing, but then some differences will be needed for different implementations.”
Kendall argues that it’s too early to commit to any particular hardware platform. Advances in sensors and other components are happening rapidly, and a truly general model should be more agnostic.
Théophile Jerbet, CEO of vertically integrated humanoid robot company Genesis AI, which raised $105 million in a seed round this year, disagrees, telling TechCrunch, “It’s too early for a brain strategy to work. Our view is there are a lot of opportunities to co-design hardware and AI.”
Gervet also touched on another hot topic in the space: how to specifically focus on the physical AI business. Robotic companies targeting specific tasks are putting their robots into the field. Grit is building solar power plants, Agility is bringing robots into industrial environments, and Bedrock is autonomously operating excavators. On the other hand, general-purpose humanoids have not left the research lab.
“No customer cares about a general-purpose robot that works with an 80% success rate,” Garbett said of the dilemma. “We’re seeing a lot of other players go general, but they don’t provide value because they’re not vertically focused. … But if you’re building a (narrow) vertical on top of GPT-2, you’re going to be crushed by companies that are building on GPT-4.”
However, the temptation to invest in specific industries is appealing because it provides not only returns but also real-world adoption data. Task-specific data may not have the diversity to advance generic models, but it is important for creating value-added robots. Bedrock CTO Kevin Peterson said his company just started with excavation as a way to understand the challenges of “operating in natural environments,” but plans to develop a layer of intelligence across its suite of construction equipment.
Managing all the data is difficult, especially with the high density of visual and lidar data. Foxglove this week announced new products built on Nvidia’s Cosmos open-weight world model. This allows engineers to use advanced natural language queries to search data and build assessments and simulations. The goal is to speed up triage and debugging, allowing model builders to iterate faster.
So what is the legendary ChatGPT moment in physical AI that Sam Altman recently said was just a few years away? Kendall points out that the world’s biggest robot deployment remains consumer cleaning robots. For him, ChatGPT’s moment will be one that excites consumers, rather than investors, who already seem pretty excited.
“One example is a car that costs less than $1,000 (hardware equivalent) and can drive itself without your eyes on it,” Kendall says. Not coincidentally, his company licenses its models to automakers to produce just that. And he sees that business as a multibillion-dollar opportunity that would allow him to build truly general-purpose, embodied AI models.
For Gervet, the moment when physical AI becomes a reality is “operations that work out of the box. You can talk to the robot in natural language and have it perform the basic tasks of the operation. For example, it can push, pull, close the laptop, clear the table, whatever you want it to do. And it works with some degree of reliability, for example, more than 80% reliable out of the box. That’s roughly ChatGPT. It’s an experience.”
Adrian Macneil, CEO of Foxglove, sees the issue a little differently.
“There will never be a ChatGPT moment in robotics,” he told TechCrunch. “What made ChatGPT instantaneous was distribution. We went from zero to 1 million active users in about a week. Real-world distribution is a lot harder than that, right? I’m very excited about the Apple II moment of robotics, the IBM PC moment of robotics. When will you be able to buy something like a home robot that starts doing useful and fun things?”
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
