Introducing TI-0.1 and TRACE: deployment-time compute scaling for robot learning
Generally intelligent robot models and hardware offer a path to scale robotics far beyond repetitive tasks in a narrow subset of the physical economy. At Tutor, we believe that deploying robots into real-world settings is a critical step toward building general robot intelligence, but robots need to be able to learn new tasks quickly and from minimal instructions to be practically useful.
TRACE uses frontier LLM agents to rebuild a robot's task and environment in code from a few Tutor demonstrations, and train reliable, human-speed policies in simulation. For a given task on a deployed robot, performance becomes a function of compute, not more real-world data. TI-0.1 is our first model trained with TRACE (see TI-0.1 running autonomously on Sonny in DF1 at 1× speed).
These results point to a clear path to deploying Sonny robots at massive scale across the physical economy. Next, we're scaling TRACE across many more tasks in our 100-robot Data Factory 1, and into real-world deployments with customers across the US.
Tutor's mission is to proliferate robots for the common good, making the physical world a meaningfully cheaper and better place to live. If you want to help, we're hiring.
Blog post, technical report, and open roles in the first comment.
Why this matters for talent
Introducing a new robot learning model, TI-0.1, alongside a deployment time compute scaling method, TRACE, is a core technical milestone for a robotics company still building toward general robot intelligence. Product announcements like this at an early stage company are usually paired with research and applied engineering hiring. Tutor Intelligence carries about 70 open roles on the jobs board, sizable for a company at this stage.
331 likes · 58 comments · 76 shares · Read the original LinkedIn post. Curated by Larry Sherwood Jr. from company announcements; analysis above is my own.