Ti0

Ti0 is Tutor Intelligence's first Vision-Language-Action (VLA) foundation model — a neural network mapping robot camera images and text instructions onto manipulator control actions. The model is targeted at dexterous bimanual manipulation in industrial settings.

Training runs on data from Data Factory 1 — a 100-robot Sonny fleet remotely supervised by Tutor Intelligence's international Tutor team. The learning loop covers behavior cloning from VR proprioceptive teleoperation (PTeleop) demonstrations, supervised fine-tuning on operator corrections, and reinforcement learning with reward shaping built from expert episode-level scores ("Tutor offline feedback").

The architecture uses velocity normalization — a custom preprocessing method that aligns the speed profiles of demonstrations across different teleoperators, reducing inconsistency in fleet-scale data. Tutor Intelligence reports that evaluating the same policy across 100 robots simultaneously allows error detection 100x faster than a classic single-robot pipeline.

Tutor Intelligence positions Ti0 as an early step toward an autonomously self-improving robot foundation model. At publication (April 2026), the model was trained on less than one week of DF1 data; the company plans to use Ti0 to power initial commercial deployments of the Sonny platform with industrial partners.