The frontier of physical AI is a Jenga game in a warehouse in San Leandro, California. Andrew Ceja, a pilot at Encord, carefully disassembles a block tower while wearing a headset that tracks his brain waves. This isn't just about collecting data; it’s about pushing the limits of what robots can learn.

Encord is one of a small but growing number of startups betting on the scarcity of real-world physical training data as the next real constraint for humanoid and warehouse robotics. They're not just managing existing data—they’re manufacturing new kinds, like brain waves and muscle signals, to train models in more nuanced ways.

Vineeth Velmurugan, Encord’s head of robot learning, sees this as a key part of the “bleeding edge” effort to solve the robotics data bottleneck. His company's work with Zander Labs aims to build an initial brain wave-tagged dataset and evaluate whether it actually improves performance before scaling up.

While tech companies scramble for physical-world data, self-driving car companies collect their own but struggle to scale. Training from video lacks real-world fidelity, making a dataset five times the size of YouTube essential—a challenge that has turned data generation into its own business. Encord is at the forefront, collecting everything from egocentric videos and remote-operated robots to brain waves and muscle signals.

For now, these pilots are just scratching the surface. But as Velmurugan notes, progress is being made, with Encord's unique visibility into programs across the industry helping them understand what works and what doesn’t. The future of physical AI may still be far off, but it’s an intriguing step in the right direction.