ML Infra, Simulation Engineering

About the Role

You will own the simulation our robots are trained and tested in. That means making it fast enough to run large-scale training, accurate enough that results transfer to hardware, and honest enough that we know which results to trust.

Simulation is only useful if it predicts reality. You will own both halves of that: the performance that makes large-scale training possible, and the fidelity that makes the results mean something.

What You'll Do

• Own our simulation stack: physics, rendering, scene generation, and the interfaces training and evaluation run against.

• Drive simulation throughput, including GPU-accelerated and massively parallel rollouts, so training is not bottlenecked by the simulator.

• Improve physics fidelity where it matters, particularly contact, friction, and compliance, and identify where it does not matter so we do not pay for accuracy we cannot use.

• Build system identification and calibration pipelines that tune simulation parameters against real hardware measurements.

• Own domain randomization strategy: what to randomize, how much, and what it costs in training time.

• Quantify and track the sim-to-real gap, so we know which results transfer and which need hardware confirmation.

• Own the authoring path for new simulation tasks, so an engineer inside or outside our team can define a task, scene, and reward without needing you.

• Build the task and scene abstractions, asset pipeline, and tooling that keep task creation cheap as the library grows.

What We're Looking For

• Built or substantially extended a physics simulator used by other people for real work.

• Deep familiarity with at least one major simulator (MuJoCo, Isaac, Drake, Bullet, or similar), including its internals and where it breaks down.

• Strong systems and performance engineering: profiling, GPU programming, and parallelism. C++ and Python.

• Working knowledge of rigid body dynamics and contact modeling, including the tradeoffs different solvers make.

• Demonstrated sim-to-real judgment: you have seen a policy that worked in simulation fail on hardware, and you know how to find out why.

• Experience building tools or frameworks other engineers use daily, with a track record of reducing the cost of a common workflow.

• Experience building tools used by people you do not sit next to, where documentation and API design matter as much as the implementation.

Bonus Points

• Experience with legged locomotion or contact-rich manipulation in simulation.

• System identification, parameter estimation, or differentiable simulation.

• Rendering and sensor simulation, including camera, depth, and tactile.

• Experience with large-scale RL training infrastructure.

• Contributions to an open source simulator or physics engine.