Tesla is building an Optimus Academy where humanoid prototypes can practice tasks and generate training data, first reported by Businessinsider. Musk says Optimus must learn by watching people and practicing tasks itself, a process he compared with how humans learn, unlike Tesla’s autonomous-driving models, which can learn from data gathered by the vehicle fleet.
The academy is being built for two related jobs: prototypes can practice a wide range of tasks and generate training data for future versions of Optimus. Musk said Tesla would need to make, in his estimate, on the order of 10,000 or 20,000 bots and have them do self-play to get the training data Optimus needs.
Robotics faces a shortage of real-world training data for teaching humanoids to navigate the physical world. Large language models such as ChatGPT, by contrast, were trained on the vast corpus of text data available on the internet.
Tesla said Wednesday that production of Optimus had begun at its Fremont factory. Musk said the company expects to begin delivering the robots to customers by the end of next year, but cautioned that production would be “somewhat bumpy” and development still had some way to go.
Musk listed durable yet responsive robotic hands and AI capable of handling open-ended tasks among Tesla’s main hurdles. Texas-based Apptronik has also set up data factories where robots load conveyor belts and sort objects to generate training data.