The physical AI sector, which applies large language model tools to robotics, is attracting substantial venture investment. However, the industry is grappling with a significant challenge: a shortage of high-quality training data for AI models, which is impacting the commercial performance of robots.
This issue was underscored by the recent performance of Unitree, a prominent Chinese robot manufacturer. After its IPO, which initially valued the company at $66 billion, its market value nearly halved. Analysts attribute this decline to the robots' current inability to perform value-creating work effectively.
At the recent Actuate conference, a gathering for developers creating AI brains for robots, the 'robotics data crisis' was a key concern. Attendees discussed the need for more diverse datasets, varied training regimes, and improved reinforcement learning scenarios to advance physical AI beyond its current 'GPT 2 era,' a reference to an earlier OpenAI model.
Companies are exploring different strategies to address these challenges. Some, like Gritt, Agility, and Bedrock, are focusing on specific tasks to deploy robots in real-world settings, such as building solar farms or operating excavators. Others, including Foxglove, are developing tools to help manage and visualize the dense visual and lidar data required for physical AI models.