Toyota Research Institute

Studying how robot experience transfers

Toyota Research Institute's robotics work includes large behavior models: policies that turn images, robot state and task instructions into physical actions. Its published research studies whether pretraining on diverse demonstrations improves later learning of particular manipulation tasks. That makes it a research contributor to robot intelligence rather than a seller of the Atlas hardware. [1]

The institute's evaluation account compares pretrained and fine-tuned models with policies trained from scratch, using physical bimanual stations and simulation. The reported benefits depend on training data and evaluation conditions; its own findings include mixed results for pretrained policies used without fine-tuning. [1]

Joint whole-body control on Atlas

The August 2025 Atlas demonstration was an actual joint project with Boston Dynamics. The technical authorship identifies both teams, and the method extends learned behavior from stationary manipulation to coordinated locomotion and manipulation. It demonstrates research on a trained robot, rather than a generally available model or household product. [2][3]