Physical AI is drawing billions in capital, but it is also confronting a fundamental bottleneck: robots still lack the high-quality training data and generalized intelligence needed to perform value-creating tasks reliably. That contrast came into sharp focus this week as Perceptron, a startup founded by former Meta research scientists, launched Isaac 0.5, an open-weight vision model designed for industrial robots, while Unitree's sharp post-IPO decline underscored investor doubts about the sector's near-term usefulness.
Perceptron was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former researchers in Meta's Fundamental AI Research division. Its new model, Isaac 0.5, aims to help robots perceive, reason, and act in warehouses and on factory floors. Unlike narrow robotics models that handle only perception or control, Isaac 0.5 combines both, and Perceptron says it was trained on a million hours of general video, first-person ego video, and UMI video of repetitive human actions. The company announced a $21 million funding round led by Bessemer Venture Partners and is targeting manufacturing, logistics, warehousing, security, mobility, and media/entertainment.
Yet the broader physical AI sector remains stuck in what some developers call the robotics data crisis. At the Actuate conference in San Francisco, organized by Foxglove and attended by about 1,500 people, infrastructure firm Avala highlighted a shortage of high-quality training data for general robot skills. Harry Mellsop, founder of simulation startup Antioch, compared the current state of physical AI to the GPT-2 era of OpenAI, before scale produced a breakthrough like ChatGPT.
Investor skepticism was visible in Unitree's market debut. The Chinese robot maker reportedly lost nearly half of its market value after a $66 billion IPO on China's STAR Market, with analysts pointing to the gap between physical capabilities and practical software know-how.
Meanwhile, autonomous vehicle expertise is being redirected into humanoids. Foxglove was founded by former Cruise employees, and both Wayve and Uber have launched robotics labs focused on humanoid form factors. Wayve CEO Alex Kendall said data infrastructure and simulation will be shared, but argued that manipulation robotics is roughly where self-driving was five years ago. Genesis AI, which raised a $105 million seed round, favors hardware-AI co-design over a general-purpose brain strategy.
The industry is also split between vertical players, such as Gritt in solar farms, Agility in industrial settings, and Bedrock in autonomous excavation, and companies pursuing general-purpose humanoids. Foxglove announced a new product built on Nvidia's Cosmos world model to help engineers query dense visual and lidar data with natural language. Whether physical AI will have a single ChatGPT moment remains contested: some leaders expect one within a few years, while Foxglove CEO Adrian Macneil argues distribution in the physical world is too hard for that analogy, and hopes instead for an Apple II moment for home robots.
