Physical AI faces challenges as sector sees both growth and setbacks

The physical AI sector, which applies large language model tools to robotics, has seen significant investment and growth, with companies like Unitree raising billions in venture capital. However, recent setbacks have emerged, including a sharp drop in Unitree's valuation following its IPO, which was initially valued at $66 billion. Analysts attribute this decline to the gap between improving robot capabilities and the ability to perform value-creating tasks. At the Actuate conference, a gathering of AI developers for robotics, the sector's momentum was evident, with attendance tripling since 2023 to reach 1500 participants. Despite this, challenges remain, particularly in the lack of high-quality training data for AI models, a problem highlighted by Avala's booth, which promised to solve the 'robotics data crisis.' Developers are seeking to address this by creating more diverse data sets, experimenting with different training regimes, and improving reinforcement learning scenarios. Harry Mellsop of Antioch suggested that physical AI is in its 'GPT 2 era,' emphasizing the need for more data and compute, especially GPUs optimized for ray tracing. Autonomous vehicles are seen as leading the way due to their ability to collect data and focus on avoiding contact. Companies like Foxglove, founded by former Cruise employees, are contributing to the tooling for model-building. Wayve's CEO, Alex Kendall, believes that investing in vehicles is a strategic move, comparing manipulation robotics to self-driving technology five years ago. He argues that while there will be shared infrastructure, specific world models for different embodiments will vary. Meanwhile, Genesis AI's CEO, Théophile Gervet, emphasized the importance of vertical focus, noting that general-purpose robots lack commercial value without specific applications. Several companies are already deploying task-specific robots, such as Gritt for solar farms, Agility in industrial settings, and Bedrock for autonomous excavators. Gervet warned that building on GPT 2 may not be competitive against GPT 4, highlighting the need for focused verticals. Foxglove recently launched a product based on Nvidia's Cosmos open-world model, enabling engineers to search data with natural language queries to speed up evaluations and simulations. Kendall envisions a 'ChatGPT moment' for physical AI as something that excites consumers, not just investors, such as eyes-off autonomy in cars for under $1000. Gervet, on the other hand, envisions a moment when robots can perform basic manipulation tasks reliably, akin to the ChatGPT experience. Adrian Macneil of Foxglove, however, doubts a direct 'ChatGPT moment' for robotics, suggesting instead a more gradual evolution similar to the Apple II or IBM PC moments. The sector continues to evolve, with companies like Wayve and Uber entering robotics labs focused on humanoid form factors. The challenge of managing dense visual and lidar data remains, but innovations in data handling and simulation are helping to address these issues as the field moves forward.