The Robot Brain Builders’ Plateau

The recent IPO flop of Unitree, China’s leading robot maker, has cast a shadow over the Physical AI sector. Analysts point to one primary reason for this downturn: robots still lack the ability to do value-creating work. This is a significant setback, but it may not be a fatal flaw in the development of truly autonomous machines.

The Actuate conference, where developers gathered to discuss AI brain-building for robots, was a testament to the sector’s enthusiasm and growth. The event has tripled in size since its inception in 2023, with over 1,500 attendees. However, beneath the surface of this excitement lies a crisis: the lack of high-quality training data for AI models.

Developers are trying to mimic the advances made by frontier AI labs by creating more diverse datasets and experimenting with different training regimes. However, this approach seems to be hitting a wall. Autonomous vehicles, while making progress, still rely heavily on data from human-driven cars. There’s a limit to how much can be achieved without significant advancements in data collection and processing.

Autonomous vehicle companies are now taking a different tack by leveraging their machine learning tooling investments to compete with dedicated humanoid makers. Tesla is experimenting with its Optimus robot, while Wayve and Uber have launched robotics labs focused on humanoid form factors. This shift highlights the increasingly blurred lines between autonomous vehicles and humanoid robots – both require significant advances in AI capabilities.

The debate rages on about what this means for the development of general-purpose humanoids. Some argue that we’re too early to commit to a single hardware platform, while others see opportunities for co-designing hardware and AI. Théophile Gervet, CEO of Genesis AI, emphasized the importance of focusing on specific tasks rather than trying to build a general-purpose robot.

The temptation to invest in a vertical is clear: it provides not just revenue but also real-world deployment data. However, this approach has its drawbacks – task-specific data may not have enough diversity to push forward general-purpose models. Robotics companies are struggling with how to balance the need for focused development with the goal of creating truly autonomous machines.

Managing all this data is another challenge altogether, especially when dealing with the sheer density of visual and lidar data. Foxglove’s new product announcement offers a glimmer of hope – its ability to search data using natural language queries could revolutionize the development process.

Alex Kendall’s example of eyes-off autonomy in cars for under $1,000 highlights the need to bring down costs and make autonomous technology accessible to a broader audience. This is an intriguing one – it shows that we’re not yet at the plateau. The excitement around Actuate shows that there’s still a lot of work to be done, and many players are eager to take on this challenge.

The stakes are high, but so is the potential reward. Addressing the fundamental issues holding us back – data collection, processing, and application – will be crucial in driving progress. The fabled ChatGPT moment may not be far off after all – and when it arrives, it will mark a turning point in human history, transforming the way we interact with machines and our surroundings forever.