Robotics investors are funding a race to build general machine intelligence, but developers still face a shortage of diverse, high-quality physical training data. Industry participants at the Actuate conference described the field as being in a “GPT-2 era”: promising enough to attract capital, yet short of the reliability and scale that would create broadly useful robots.
The conference drew 1,500 attendees, three times its size when it began in 2023, according to organizer Foxglove. The growth reflects excitement, while exhibitor promises to solve a “robotics data crisis” reveal the technical bottleneck beneath it.
Physical AI cannot learn from the web alone
Language models can train on vast collections of digital text. Robots need data about perception, motion, forces, objects and failure in real environments. General systems remain far from handling arbitrary tasks, while task-specific end-to-end learning has not consistently delivered commercial reliability.
Harry Mellsop, founder of simulation startup Antioch, uses the GPT-2 comparison to argue that more data and compute are required. High-fidelity simulation also depends on graphics processors optimized for ray tracing, creating another infrastructure cost.
Autonomous vehicles are further ahead because human-driven cars can collect relevant road data and because avoiding contact is narrower than manipulating varied objects. Tools created for vehicle development are now moving into other robotics domains. Foxglove, for example, was founded by former Cruise employees.
This does not prove that methods transfer unchanged. A driving simulator and a model controlling hands or construction equipment need different post-training and representations of the world.
General brains compete with vertical deployments
Wayve CEO Alex Kendall expects data infrastructure, simulation and machine-learning operations to be shared across vehicles and other robots, while the world model changes for each physical form. Wayve and Uber have both launched humanoid robotics laboratories connected to their vehicle work.
Genesis AI CEO Théophile Gervet takes a more integrated view. He argues that the industry is too early for a brain-only strategy and sees opportunities to co-design hardware and intelligence. He also notes a commercial dilemma: narrowly focused robots can reach customers, but a company specializing too early may be overtaken by a more capable general model.
Several task-specific systems are already operating: Gritt builds solar farms, Agility deploys industrial robots and Bedrock automates excavators. Bedrock says excavation is a starting point for learning “manipulation in the wild,” with an intelligence layer eventually spanning additional construction machines.
Field deployment brings revenue and real data, though information from one task may lack the diversity needed for a general model. The source provides no shared benchmark that settles whether specialization or generality is advancing faster.
Robotics may get a PC moment, not a ChatGPT moment
Foxglove announced a tool using Nvidia’s open Cosmos world model to search dense visual and lidar records with natural-language queries. The aim is to speed debugging, evaluation and simulation by helping engineers locate relevant events. It addresses data management rather than the underlying absence of enough varied experience.
Industry leaders also disagree about the milestone that would signal a breakthrough. Kendall points to inexpensive eyes-off vehicle autonomy as one possible consumer-scale event. Gervet imagines manipulation that works above 80 percent reliability from ordinary language commands.
Foxglove CEO Adrian Macneil rejects the premise of one ChatGPT moment because physical distribution is slower than software adoption. He would instead welcome an Apple II or IBM PC stage: a home robot that consumers can buy and use for genuinely helpful tasks.
The disagreement is informative. Capital and demonstrations can move quickly, while collecting data, manufacturing hardware and deploying machines cannot. Physical AI is advancing, but the next stage depends on infrastructure and field reliability more than a single viral interface.
