CES® 2027 NEWS
ROAD TO CES 2027Live now
ROAD TO CES · AN INTELITOYS EVENTJAN 6–9 · VEGAS
Home Robots

Tutor Intelligence Says Its TRACE Agents Trained TI-0.1 in Simulation From a Handful of Demonstrations

The robot maker argues compute, not real-world trial and error, is now the limiting resource for teaching machines new jobs.. Tutor Intelligence introduced TI-0.1, the first model in its TI series, built with an agentic harness called TRACE that reconstructs a robot task in simulation and trains a policy there.

InteliToys Editors ? Tutor Intelligence TI-0.1
Sonny, Tutor Intelligence's two-armed mobile robot worker, standing in a warehouse aisle between shelves (image: Tutor Intelligence)

Tutor Intelligence used a blog post to introduce TI-0.1, the opening entry in its TI line of robot models, together with TRACE, an agentic harness the company says lets frontier language models reconstruct a robot's task in simulation and train a policy there. The pitch is that a machine can be taught a new job with only a few human demonstrations instead of months of engineering. For a company selling robot labor rather than robot hardware, that reframing is the real product news: the model is only part of what Tutor is shipping.

The company frames the problem as a data shortage. Tutor says robots still mostly run narrow routines in structured settings, and that nine out of ten businesses in manufacturing and logistics do not operate a single robot. Its answer is to seed a simulation with a few demonstrations recorded by human operators, then let agents refine the task there.

Tutor says TRACE, which it spells out as Telemetry-grounded Reconstruction, Alignment, and Closed-loop Evaluation, runs in three stages. Agents first rebuild the scene and the motion the robot recorded, replaying the logged commands so the physics can be checked rather than just the picture. A policy then rehearses the job inside that simulation with reinforcement learning. Finally the agents review the rollouts and adjust the task until the behavior matches what was intended, work the company says once consumed weeks of scene building, physics tuning and reward debugging.

Ride the Las Vegas Monorail to CES - stops at the Las Vegas Convention Center

Tutor says the payoff is speed and data efficiency. Starting from just a few demonstrations, TI-0.1 learns its task entirely in simulation and then runs autonomously on the company's Sonny robot at human speed, and the company posted video it describes as the model running at 1x speed. Every real-world trial exposes where the simulator still diverges from reality, and the agents feed that back into the next training round.

Tutor also rolled out Golem, a robot software stack the company says records what each machine saw, what it was asked to do and how it moved on one shared timeline. Camera handling and arm control run in Rust, with policy inference handled by the robot's own GPU. Compared with its earlier Python-based stack, Tutor says Golem used 90% less CPU and 79% less RAM, with 68% lower camera-streaming latency. Tutor says that shared timeline is what makes the loop practical, because a policy can be inspected against the exact footage and joint readings that produced it.

The company is candid that this is an early step. Tutor says it is now scaling TRACE across more tasks at its DF1 facility and into deployments with customers across the United States, and it has not published the raw benchmarks behind the Golem comparisons or the success rates that would let a buyer judge how far the simulation results carry into a real facility. What remains unknown is how many demonstrations a genuinely new task needs, how often the simulator has to be corrected by hand, and whether the human-speed claim holds up in facilities that are less orderly than a demonstration aisle.

Source: Tutor Intelligence.

InteliToys is editorial. We do not sell Tutor Intelligence TI-0.1. Watch the source tape in the branded lightbox below.