Double turns what skilled professionals know into training and evaluation data for robots. An expert defines a task and the standard the result has to meet. We build that task in simulation. Experts record demonstrations of it, and correct a robot's failures if the team already has a policy. The robotics team trains its own model on the data and tests it on reserved scenarios, judged against the expert's criteria.
Why experts
Much of skilled work was never written down. It lives in the people who do the job: the order of operations, what counts as done, and how to recover from a mistake. A mechanic feels for cross-threading before it binds. An electrician listens for the click, then does a pull test. Those details belong in the training, and the people who know the work are the ones who can supply them.
The experts are people with practical experience in a trade or technical field who can explain the steps, checks and common mistakes in their work. They don't need a robotics background. Each project provides the instructions and controls needed for the task.
Why simulation
Simulation lets the same task run under many variations, and lets an expert pick up from the exact state where a robot went wrong. That is hard to do repeatedly on real hardware.
Step 1: Define the skill
An expert writes down the procedure, the situations it has to handle and the quality bar. We build it in simulation: the environment, the scenarios, the checks and reference demonstrations. The task runs in Isaac Lab, MuJoCo, ManiSkill, Genesis or the team's own simulator, with task-specific objects, materials and settings where needed.
The robotics team then signs off on the task, including anything the simulation can't represent. More on task creation.
Step 2: Capture demonstrations
Experts record the technique, the actions and notes on the decisions they make. A team doesn't need a policy to start. A recording can cover the whole task or a single subskill, and each one comes with annotations of the decisions, conditions and quality behind it.
- Practice first. Every expert learns the teleoperation controls on practice tasks.
- Same view as the robot. Experts see the same cameras and state the policy trains on, plus force where the simulation models it.
- Say why, as it happens. Experts mark stages and note their decisions and the result.
- Reviewed before delivery. Only reviewed episodes are delivered.
Variations are agreed up front, so a few sessions can cover many situations, and every episode is labelled by where it came from. Data is delivered as LeRobot, RLDS, HDF5 or the team's own format, with usage rights agreed before work starts. More on training data.
Step 3: Correct the failures
If a team shares a checkpoint, we run it in the same simulation. When an attempt fails, an expert takes over from the exact state where it went wrong and recovers it. Those interventions show the policy what the right response looks like in the situations it actually gets wrong. More on corrections.
Step 4: Evaluate
The test is agreed before anyone trains for it. The rubric and test scenarios are set with the team before any data is collected, and the test scenarios never appear in the training data.
- Experts define what correct means. Each criterion is checked automatically or by review.
- Every policy version runs on the same reserved scenarios.
- Results are reported by criterion and scenario, across versions, so you can see where it breaks down.
We run the evaluation and report, or package it for the team to run in-house. Expert labels and reward design are available as separate work. More on evaluation.
What stays with the robotics team
The team keeps its model and training pipeline. It shares the skill it needs and a checkpoint if it has one. We build the environment and return recordings, annotations and evaluation results. Controls, observations and format are agreed up front, and we recruit, vet and pay every expert.
How a pilot runs
- Scope the task. The team brings the skill, and failure examples if it has a policy. We scope the environment and the first deliverable.
- Set up a sample session. Confirm robot controls, observations and task checks, and inspect a sample recording.
- Run and collect. Experts record demonstrations. With a checkpoint, we also run it and experts correct its failures.
- Train and evaluate. The team trains its model and compares the result on the reserved scenarios.
- Decide what comes next. Extend the work once the pilot shows a useful result and a repeatable process.
See Double for robotics teams for all four services.
For experts
Experts contribute in three ways: explaining a task, reviewing a robot's attempt, and demonstrating or correcting the work with the controls provided. Most of it happens from a computer. Project work is paid, with scope, time and payment terms agreed before it starts. If your field is listed, see the open roles. If it isn't, you can still apply.