Multitask AI models help robots tackle complex tasks with improved efficiency

A robot assembling a breakfast tray with the precision of a seasoned chef might sound futuristic, but researchers at the Toyota Research Institute have made it reality using artificial intelligence that learns like humans do—by watching and imitating. Their breakthrough involves training robots on massive datasets of nearly 1,700 hours of demonstrations across more than 500 different tasks, from slicing apples to installing bicycle brake rotors.

These "large behavior models" proved remarkably efficient when fine-tuned for specific jobs, achieving the same results with three to five times less training data than traditional approaches. The efficiency matters because collecting robot training data is expensive and time-consuming.

"Our findings largely support the recent surge in popularity of LBM-style robot foundation models, adding to evidence that large-scale pretraining on diverse robot data is a viable path toward more capable robots," said Jose Barreiros, a TRI researcher.

After 1,800 real-world trials, the multitask-trained robots showed superior adaptability when encountering unfamiliar situations, suggesting a future where robots might learn new skills as naturally as experienced workers picking up related tasks.