Unitree, China's leading robot maker, saw its valuation drop by nearly half following its $66 billion IPO, signaling growing concerns in the physical AI sector. Analysts attribute the decline to the lack of high-quality training data for AI models, which remains a critical barrier to commercial success. While physical capabilities of robots are improving, their ability to perform value-creating tasks is still limited, according to industry observers.
At last week’s Actuate conference, a gathering of developers building AI brains for robots, the event had tripled in size since 2023, with 1,500 attendees, according to Foxglove, the organizer. The conference also highlighted the industry’s struggle with the robotics data crisis, as companies like Avala promised to solve it. Developers are turning to better datasets, varied training regimes, and improved reinforcement learning scenarios to overcome these challenges. Harry Mellsop, a founder of Antioch, described physical AI as being in its 'GPT-2 era,' suggesting more data and compute, particularly GPUs optimized for ray tracing, will be needed to advance the field.
The physical AI sector is facing a significant challenge in creating generalized robots capable of performing any task, with autonomous vehicles leading the way due to their ability to collect relevant data and avoid contact. Many of the tools used in model-building come from autonomous vehicle companies, such as Foxglove, which was founded by former Cruise employees. Now, these car companies are increasingly investing in ML tooling to compete with dedicated humanoid makers. Tesla is already pursuing this with its Optimus robot, and both Wayve and Uber have launched robotics labs focused on humanoid form factors.
Source: techcrunch
