The rapid advancement of artificial intelligence has largely followed the 'scaling hypothesis,' which suggests that increasing model size, compute, and training data leads to predictable gains in intelligence. While this has worked for language models—which can be trained on trillions of tokens of existing internet text—robotics faces a different challenge. There is no equivalent public stock of high-quality data for physical tasks, such as folding laundry or manipulating objects. Because robot training requires examples of actions performed correctly (or a way to score an attempt), the field is currently grappling with how to generate the necessary volume of data to move beyond simple, repetitive tasks.
One primary method for gathering this data is teleoperation, where a human controls a robot's movements via a remote interface. This allows for 'behavioral cloning,' where a model learns to mimic the expert's actions. While this produces data that closely matches the robot's actual physical capabilities, it is difficult to scale due to the high cost of human labor and the lack of tactile feedback for the operator. Alternatively, researchers use simulation to create 'synthetic' data. Simulation allows for thousands of robots to be trained simultaneously in a digital environment, which is significantly faster and cheaper than real-world testing. However, this often leads to the 'sim-to-real gap,' where a model performs well in a perfect digital world but fails in reality because the simulation cannot perfectly model complex physics like friction or the behavior of soft, deformable objects.
To mitigate these limitations, companies are turning to human-centric data collection. This includes 'egocentric' video, where people wear head-mounted cameras to record their own movements, which are then mapped to robot actions. Other methods involve 'Universal Manipulation Interfaces' (UMI), where humans wear specialized manipulators to perform tasks, providing high-fidelity data for robots with similar hardware. While these methods are more scalable than one-on-one teleoperation, they still face hurdles in capturing the nuanced sensory feedback—such as the precise pressure and texture of an object—that humans naturally process through their nervous systems. For now, the path to general-purpose robotics relies on a mix of these strategies to build a 'data flywheel' where real-world deployments eventually feed back into the training loop to improve future models.