Build with us.

About the role

We want the robot to learn connector work from demonstrations, touch and its own test results. You’ll find out where learning helps and what it takes to switch to a new connector.

You’ll own data, training and running models on the robot, with the founders and the controls engineer. A result counts when the joint passes its test, the robot recovers on its own and no one has to step in. Start date follows the prototype build.

Pay

$150,000–$180,000 USD per year + equity.

Responsibilities

• Build tools that record demonstrations and corrections, in sync with what the robot did and whether the part passed.
• Train imitation-learning policies. Try reinforcement learning where it fixes a real failure.
• Run models on the robot within its calibration, latency and control limits.
• Version data and checkpoints. Build tools to replay attempts and compare runs.
• Test policies against hand-written control on fresh parts: misses, damage, manual fixes.
• Move to a second connector and count what it took: demos, tooling and engineering time.

Experience

• You’ve trained a policy, run it on a real robot and dug into why it failed.
• Training and evaluation code in Python, with PyTorch or JAX.
• Hands-on imitation or reinforcement learning, and how the data you collect shapes the result.
• You know robot observations, actions and timing, and can debug everything around the model.
• You can set up a test that tells a better model apart from a change in parts or fixtures.

Useful experience

• Insertion tasks, action-chunking policies or learning from human corrections.
• Force or tactile sensing, LeRobot or SERL.
• C++, ROS 2, edge inference or adapting policies across parts and fixtures.

First project

Collect demonstrations and build a repeatable test for the first connector. Train a first policy, compare it with the programmed controller, and let its failures pick the next experiment.