The article discusses the current state and challenges of the physical AI sector, which has attracted significant venture capital to apply large language model (LLM) techniques to robotics. Despite the hype, exemplified by Unitree’s $66 billion valuation post-IPO, the sector faces a critical issue: robots’ physical capabilities are advancing, but they still lack the practical know-how to perform value-creating work. The Actuate conference, a gathering of developers building AI brains for robots, highlighted both the sector’s excitement and its risks, notably the “robotics data crisis” – the scarcity of high-quality training data for AI models. Attempts to create generalized robots capable of any task remain elusive, and task-specific end-to-end learning has yet to yield commercially reliable products.
Developers are responding by emulating the strategies of frontier AI labs: expanding and diversifying datasets, experimenting with training regimes, and refining reinforcement learning scenarios. Harry Mellsop of Antioch analogizes physical AI’s current stage to the pre-ChatGPT era of OpenAI’s GPT-2, suggesting that substantial increases in data and computational power, particularly GPUs optimized for ray tracing, are necessary to overcome existing hurdles. The article notes that autonomous vehicles are relatively ahead due to accessible data collection and a primary task focused on collision avoidance rather than complex manipulation. Consequently, companies like Tesla, Wayve, and Uber are leveraging their automotive AI tooling investments to venture into humanoid robotics, despite differing views on whether to prioritize hardware co-design or maintain hardware-agnostic AI models. The debate also touches on the strategic focus of robotics companies, with some targeting specific tasks and real-world deployments while others pursue general-purpose humanoids, each approach presenting distinct challenges and opportunities.