Physical AI — applying techniques behind large language models to robotics — has attracted substantial venture capital, with startups and companies raising billions. That enthusiasm helped drive a large IPO for Unitree, China’s leading robot maker, which reached a $66 billion valuation after its public debut. This week the stock collapsed, wiping out nearly half the company’s value. Analysts point to a clear issue: while robots’ physical abilities are improving, they still lack the know-how to perform reliably valuable work.
Developer community and the data crisis
At the Actuate conference, where engineers and firms building AI "brains" for robots gathered, excitement was palpable. The event has tripled in size since its 2023 launch and drew about 1,500 attendees, according to organizer Foxglove. But the risks are visible too: a booth from Avala promised to solve “the robotics data crisis.”
That crisis refers to a shortage of high-quality, diverse training datasets for physical AI. General-purpose robots capable of arbitrary tasks remain far off, and end-to-end learning approaches have yet to produce commercially reliable products even for specific tasks. Developers say the remedy is to emulate frontier AI labs: find or generate more diverse data, experiment with training regimes, and design better reinforcement learning scenarios.
Harry Mellsop, founder of Antioch, a startup building simulation tools for model developers, describes physical AI as being in its “GPT‑2 era” — a stage preceding a broad consumer breakthrough. He argues more data and compute power are needed to clear the next hump, notably GPUs optimized for ray tracing used to produce high-fidelity simulations.
Autonomous vehicles, shared tooling, and vertical focus
Autonomous vehicles are furthest ahead partly because it’s easier to collect relevant data from human-driven cars and because the main task is avoiding contact rather than manipulating complex physical objects. Much tooling for model building has emerged from AV companies; Foxglove, for example, was founded by former employees of Cruise, General Motors’ previous self-driving effort.
Automakers increasingly believe investments in ML tooling will let them compete with dedicated humanoid makers. Tesla’s Optimus is one example; AV-focused Wayve and ride-hail giant Uber have launched humanoid robotics R&D labs.
Alex Kendall, CEO of Wayve, says starting in vehicles makes sense: “manipulation robotics is like self-driving five years ago.” He expects data infrastructure, simulation and MLOps to be largely shared, though each embodiment will need different post-training world models in the simulator. Kendall also cautions it’s premature to commit to a single hardware platform because sensors and components are evolving quickly.
Théophile Gervet, CEO of Genesis AI — a vertically integrated humanoid robotics company that raised a $105 million seed round this year — disagrees on the hardware-agnostic point, arguing there are many opportunities to co-design hardware and AI now.
Gervet also emphasized the strategic choice many companies face: focusing on specific tasks or building general-purpose humanoids. Task-focused firms are shipping products: Gritt builds solar farm solutions, Agility deploys robots in industrial settings, and Bedrock operates excavators autonomously. General-purpose humanoids largely remain in labs.
“No customer cares about the general purpose robot that works at 80% success rate,” Gervet said, noting that vertical focus brings revenue and crucial deployment data. He warned, however, that startups building vertically on earlier-generation base models risk being outcompeted by firms leveraging later, stronger base models.
Bedrock CTO Kevin Peterson said his company started with excavation to understand the challenges of “manipulation in the wild,” and plans to develop an intelligence layer that spans multiple construction machines.
Data management, simulation tools and searching for the breakout
Managing dense visual and lidar data is challenging. This week Foxglove announced a product built on Nvidia’s Cosmos open-weight world model that enables engineers to search datasets with sophisticated natural-language queries and rapidly assemble evaluations and simulations. The goal is faster triage and debugging so model builders can iterate more quickly.
What will be the physical AI equivalent of ChatGPT — the consumer-facing moment Sam Altman said is a few years away? Kendall argues the breakthrough that matters is one that excites consumers, not just investors. Today the largest robot deployments are consumer vacuum robots. For him, a ChatGPT-like moment would be a leap such as achieving eyes-off autonomy in a car with less than $1,000 of hardware — and Wayve is licensing models to automakers toward that end. That business, which he sees as a multibillion-dollar opportunity, could enable building a truly general embodied AI model.
Gervet defines the tipping point as “manipulation that just works out of the box”: being able to tell a robot in natural language to perform basic manipulation tasks (push, pull, close a laptop, clean a table) with roughly 80%+ out-of-the-box reliability — a comparable consumer experience to ChatGPT.
Adrian Macneil, CEO of Foxglove, disagrees that there will be a single ChatGPT moment for robotics. He notes ChatGPT’s impact was driven by rapid distribution — going from zero to a million active users in about a week — and says real-world distribution is far harder for robots. Instead, he would welcome an “Apple II moment” or “IBM PC moment” for robotics: the point when consumers can buy a home robot that starts doing useful and enjoyable tasks.
Conclusion
The field of physical AI is fueled by capital and experimentation, but progress toward general-purpose, commercially reliable robots is constrained by a lack of diverse, high-quality data, the need for better simulation hardware and tooling, and the value of real-world deployment data—often obtained by targeting narrow verticals. Industry leaders agree that shared tooling, improved data management and continued iteration across multiple approaches (vehicle-first, vertical focus, or hardware–AI co-design) will be necessary to achieve broader breakthroughs.
