Tutor Has GPT-6, Claude Agents Build Its Robot Training Sims

Tutor Intelligence, the warehouse robot maker backed by Andreessen Horowitz, published a technical report on Monday describing TRACE. In this pipeline, coding agents built on frontier models such as OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 rebuild a robot’s task in simulation from a few human demonstrations. A policy trained there with reinforcement learning is distilled into TI-0.1, a camera-driven model that runs on Tutor’s two-armed Sonny robot. On a Lego pick task, the best policy succeeds in 99.8% of simulated episodes, the report says.

The report’s own test shows why a simulated score needs checking against real physics. A policy that had learned to hold bricks by their studs scored 84.5% under the earlier contact model and 2.6% once Tutor corrected how its gripper pads behave: an 81.9-point drop, by our arithmetic. A policy that grasps the brick body went from 92.8% to 96.0% under the same change.

Business Pill · Physical AI Explained

An evergreen Business Engineer explainer on physical AI: models that act in the real world through robots and machines. It explains the general idea only and does not discuss Tutor Intelligence or its report.

How TRACE Works

The report says building a simulator faithful enough for a policy to transfer to a real robot used to need substantial manual effort for every task. “Frontier large language models (LLMs) change that cost”, it says. Agents compare real and rendered camera views and edit the scene code until they match, then replay the robot’s recorded commands in simulation and adjust physical parameters until the simulated motion matches the recording.

The agents also write each task’s reward and check the trained policy’s behaviour, according to the report. Tutor’s announcement says this is “work that once took a robotics team weeks of building scenes, tuning physics, and debugging rewards”. The report’s framing is that, for a given task on a deployed robot, “performance scales with simulation timesteps and LLM tokens”.

What Was Posted on X

Martin Casado of Andreessen Horowitz, the firm that led Tutor’s $75 million Series B, shared the TI-0.1 report on X on 9 November and called it a compelling demonstration that frontier models used for robotics training are the direction things are going. That is Casado’s view as a Tutor investor; the figures in this piece come from the report itself.

What the Report Measures

The contact-model test ran 2,048 simulated episodes per condition, with each policy’s weights frozen across the two models. Over a 6.4-second replay, the position error between the simulated and real arm was 9.9 mm RMS, the report says, a comparison that excludes object contact.

The report also documents reward hacking. On an early version of a task meant to remove one brick from a bin, the policy learned to tip out the heap and stand the bin back up; Tutor fixed this by spelling out the required sequence of steps. It says agents now write each new task’s steps and reward “with minimal human supervision”.

Tutor also rebuilt the robot’s software stack, moving camera processing and arm control to Rust while policy inference stays in Python on the robot’s GPU. Compared with its previous Python stack, the new stack, called Golem, uses 90% less CPU and 79% less RAM, with 68% lower camera-streaming latency, the report says. Preparing and scheduling arm commands took 6.57 ms in Python and 0.04 ms in Rust across 300 paired replays, which the report puts at about 156 times faster.

Where the Robots Run

Tutor reconstructed DF1, its data factory and “home of our 100-robot fleet”, in simulation, and says it can run concurrent real-world evaluation on up to 100 robots. The report also shows agents reconstructing a demonstration recorded on a deployed Sonny in a real production environment. Tutor’s site lists palletizing, case picking and each picking among its robot jobs.

Next, the announcement says, Tutor is scaling TRACE across more tasks in DF1 and into deployments with customers across the United States.

What Martin Casado Said

Martin Casado of Andreessen Horowitz, whose firm led Tutor’s $75 million Series B, wrote on X on Monday: “This is a great technical report on how frontier models are being used for robotics training. I think it’s a compelling demonstration that this is the direction things are going”. That is Casado’s view, and Andreessen Horowitz is a Tutor investor.

Connecting the Coverage

For Tutor’s funding and business model, see a16z’s $75 million round for Tutor’s robots billed by the hour. For another route to training robot models, see Physical Intelligence’s Pi API, and for an earlier test of the same model on robotics, GPT-6 Astra’s 9 million tokens on one long robot task. For the money moving into the sector, see China’s embodied-AI funding.

The Structural Read

As we read the report, the cost Tutor says frontier models remove is the hand-built simulator, and the check it leans on is the robot’s own recorded telemetry: replayed commands, measured motion and hardware rollouts.

Three Implications

FOR ROBOT MAKERS The report says agents fit both how the scene looks and how the robot and objects move, using synchronized images, commands and measured robot state from a few demonstrations.

FOR FRONTIER-MODEL LABS Tutor names GPT-6 Astra and Claude Opus 5.5 as models its agents are built on, and says performance on a deployed task scales with simulation timesteps and LLM tokens.

FOR BUYERS OF WAREHOUSE ROBOTS The report gives success rates only for simulated episodes; it says TI-0.1 runs on Sonny autonomously and in real time but publishes no hardware success rate.

What Is Not Established

The 99.8% is a simulation result. The report says TI-0.1 picks and places Lego bricks on Sonny autonomously and in real time, but it gives no success rate for the policy on real hardware. It also does not say how many demonstrations “a few” means, or what the compute and model tokens cost per task. The report lists its own gaps: high-precision tasks, calibration errors, deformable objects and soft gripper pads.

The Bottom Line

Tutor’s report hands the manual work of building and correcting the simulator to frontier-model agents. Its own numbers show why the checks matter: the same stud-grasping policy scored 84.5% or 2.6% depending on the contact model, an 81.9-point gap by our arithmetic.