Developers building artificial intelligence for robots are grappling with a severe shortage of high-quality training data, even as investment pours into the sector and companies raise billions to bring large language model techniques to physical machines. That challenge became starkly visible this week when Unitree, China's top robot manufacturer, saw its valuation collapse by nearly half — dropping from $66 billion after its recent IPO on China's tech-heavy stock exchange — because analysts say the robots still can't perform work that creates real value, according to a TechCrunch report published August 26. The excitement around physical AI remains intense, but the industry faces what developers call a "robotics data crisis" that's keeping the technology stuck in its early stages.
Last week's Actuate conference, where developers working on AI brains for robots gathered, drew 1,500 attendees — triple the size since launching in 2023, organizer Foxglove reported. The event showcased both the sector's growth and its fundamental problem: a booth for physical AI infrastructure company Avala advertised a solution to "the robotics data crisis." Efforts to create generalized robots capable of any task remain distant goals, and end-to-end learning for specific tasks hasn't yet produced products with dependable commercial performance. Autonomous vehicles have made the most progress, partly because developers can gather relevant data from human-driven cars and partly because the primary objective is avoiding collisions rather than physically manipulating objects. Tesla is already pursuing this crossover with its Optimus robot, while both autonomous vehicle startup Wayve and rideshare giant Uber have opened robotics labs focused on humanoid forms as research and development projects. Companies targeting narrow tasks — like Gritt building solar farms, Agility deploying robots in factories, and Bedrock operating excavators without human operators — are getting machines into the field, while general-purpose humanoids remain confined to laboratories.
Harry Mellsop, founder of simulation tooling startup Antioch, told the publication that physical AI is currently in its "GPT-2 era," referencing the OpenAI model that came before ChatGPT arrived. Developers need more diverse datasets, different training approaches, and improved reinforcement learning scenarios to advance, particularly GPUs optimized for ray tracing to build high-fidelity simulations. Théophile Gervet, CEO of Genesis AI, a vertically integrated humanoid robotics company that raised a $105 million seed round this year, described the sector's dilemma: "No customer cares about the general-purpose robot that works at 80% success rate," he said, adding that companies focused on general applications provide no value without vertical specialization, but building for a narrow vertical on today's primitive models means getting crushed when competitors deploy more advanced AI. Foxglove this week launched a new product built on Nvidia's Cosmos open-weight world model, allowing engineers to search dense visual and lidar data using sophisticated natural language queries for faster debugging and iteration.
The industry's "ChatGPT moment" remains undefined, with different leaders offering conflicting visions. Alex Kendall, CEO of Wayve, pointed out that consumer vacuum robots still represent the world's largest robot deployment and argued a breakthrough would excite consumers, not just investors — such as eyes-off autonomy for under $1,000 in hardware cost for cars, which his company is pursuing by licensing models to automakers. Gervet defined the milestone as manipulation that works immediately out of the box, where users can speak to a robot in natural language and have it complete any basic task like pushing, pulling, or cleaning with at least 80% reliability. Adrian Macneil, Foxglove's CEO, took a sharper view, saying "There will not be a ChatGPT moment for robotics" because real-world distribution is far harder than software's instant scale, and instead he'd welcome an "Apple II moment" when consumers can buy a home robot that does useful and enjoyable tasks. The report suggests the path forward requires either massive increases in data and computing power for general models, or acceptance that near-term commercial success demands narrow specialization even as the industry builds toward broader capabilities. The tension between generating revenue through vertical focus and developing the diverse datasets needed for truly general intelligence remains the sector's central strategic challenge, one that appears likely to persist until breakthrough improvements in model performance make reliable, adaptable robots economically viable across multiple use cases.