Our recordings start in a room full of sensors and are meant to end in a robot’s policy. We need the engineer who owns everything in between. Experience with reinforcement learning is preferred.
The work
- Own the recording software: sensor drivers, capture, and time sync across cameras, eye trackers and wearables.
- Build the pipeline from capture to dataset: ingest, encrypted storage and processing jobs at scale.
- Export clean episodes in MCAP, LeRobot v3, RLDS and HDF5, and check that they load on the other side.
- Show that the data works: train and evaluate policies on it, from imitation learning to reinforcement learning, simulation first.
- Build the web tools the team runs on: session tracking, quality review and labelling.
- Work with collaborating labs on formats, schemas and what their models need.
You bring
- Two or more years shipping software across the stack, with robots or machine learning in the mix.
- Strong Python, plus C++ or TypeScript.
- ROS 2, sensor data and time synchronisation.
- PyTorch: you have trained models and read the loss curve honestly.
- A backend and a frontend you built and kept running (APIs, Postgres, React or similar).
- Linux, Docker and the cloud.
Preferred
- Hands-on reinforcement learning: PPO, SAC or similar, reward design, sim-to-real.
- Isaac Sim or Isaac Lab, MuJoCo, or similar simulators.
- LeRobot, robot arms or embedded hardware.