Robot AI Push Runs Into Training Data Bottleneck

Physical AI developers are adapting the data, compute and reinforcement-learning techniques behind large language models to machines that must perceive and manipulate the real world. Yet general-purpose robots remain far from commercial readiness. Antioch co-founder Harry Mellsop described the sector as being in its “GPT-2 era,” with progress constrained by scarce, diverse training data and the heavy computing demands of high-fidelity simulation.

TechCrunch reported on Aug. 26, 2026, that Foxglove’s Actuate conference drew 1,500 attendees last week, triple its size since launching in 2023. Genesis AI raised a $105 million seed round this year, while Unitree reached a $66 billion valuation after going public before losing nearly half its value this week. The reversal underscored the gap between improving robotic hardware and systems that can perform reliable, commercially valuable work.