Your starting point
The transition is not simply replacing a text dataset with robot videos. A policy produces actions that affect its next observations. Choose a narrow manipulation or control task and make the evaluation conditions explicit. Before targeting a vacancy, distinguish policy research, data infrastructure, simulation engineering and deployment: they can belong to the same team but need different evidence.
What you can carry over
- Experiment design
- Carry over baselines, controlled comparisons and clear evaluation splits.
- Data pipelines
- Extend them to synchronized observations, actions and robot state.
- Model deployment
- Investigate inference timing and what happens when observations or actions are delayed.
Skills to build next
Understand the action representation, observation timing and the distinction between an offline metric and success over a complete task. Learn how demonstrations are generated and which conditions they omit. Familiarize yourself with coordinate frames and the robot’s operational limits before treating a policy as a deployable component. A strong predictive metric does not by itself establish reliable closed-loop behavior.
Portfolio project: Evaluate an imitation-learning baseline
This is a suggested project, not a required credential or a promise of employability.
- Choose a documented dataset and supported task. Record the dataset version, task definition and what counts as success before training.
- Create a baseline and an evaluation split that is meaningfully separate from training demonstrations. State whether separation is by episode, environment, object or another factor.
- Run repeated evaluations under a declared set of conditions. Preserve failures, report the number of trials and compare one deliberate change against the baseline.
- Produce a model card-style project note covering data, action space, compute, results and limitations. If you only evaluated in simulation, say so prominently.
Keep the evidence honest: Use a simulator or an appropriately supervised lab setup. A learned policy and a visually successful demo are not a safety validation for a physical robot.
What to put in the evidence package
Include both the model work and the engineering that makes the result inspectable: configuration, data preparation, evaluation scripts and unsuccessful trials. Explain why you chose the comparison and what alternative explanation could account for the result. If a benchmark or tutorial supplies most of the implementation, identify your own contribution precisely instead of presenting the entire system as original work.
Turn the project into a focused application
Use research requirements to choose the right application route, not as a universal rule about robotics. Some vacancies may ask for a research record; others emphasize engineering ownership. This guide does not measure the share requiring a PhD. Where your experience is strongest in infrastructure, look for that responsibility explicitly rather than assuming every robot-learning opening is a policy-science role.
Questions to ask the employer
- Would I own policy research, data collection, evaluation or deployment?
- What fraction of the work uses real robots rather than offline data or simulation?
- Does this particular role require publications, a doctorate or a specific hardware background?
Learning references
Start with the relevant section of these primary resources. They support the technical learning topics; the project plan and career advice above are our editorial suggestions, not employer endorsements.
- LeRobot documentation — A reference for recording demonstrations, training policies, deployment and simulation-based evaluation.
- MoveIt tutorials — A complementary introduction to motion planning and pick-and-place, distinct from learning a policy.