Simulation & Robot Learning Engineer

You'll own the pipeline from environment to deployed behavior: authoring scenes and tasks, training policies (RL, imitation/diffusion, VLAs, or world models), scaling data collection and generation, and using agentic AI to automate the slow parts- scene and task generation, reward design, domain randomization, and evaluation- so the team iterates in hours, not weeks. You'll work across hard manipulation regimes, from rigid and articulated bodies to deformables (cloth, cable, soft objects), on platforms ranging from single cobot arms and bimanual setups to whole-body and wheeled humanoids.

What You'll Do

• Build and maintain high-fidelity simulation environments (Isaac Sim/Lab, MuJoCo, Drake, Gazebo, SAPIEN, or similar), scenes, assets, physics/contact tuning, and task definitions for manipulation.
• Train deployable robot policies: reinforcement learning, imitation/diffusion policies, VLAs, and/or world models, and drive them from sim to robust real-world behavior.
• Close the sim-to-real gap: domain randomization, system identification, real-to-sim calibration, and the messy debugging that makes a policy survive contact with reality.
• Scale data: build pipelines for large-scale synthetic data generation, teleop/demonstration data, procedural task and scene generation, and automated curricula.
• Use agentic AI to expedite the whole loop: auto-generating environments, tasks, reward functions, randomization configs, and evaluations, and turn slow manual work into fast automated iteration.
• Handle hard manipulation regimes: rigid, articulated, and deformable objects (cloth, cable, soft materials), and contact-rich tasks.
• Build evaluation and benchmarking infrastructure so that performance in sim is a trustworthy predictor of performance on hardware.
• Work closely with controls, perception, mechanical, and the rest of the learning team to turn a target behavior into smooth, safe, repeatable arm and whole-body motion on the robot.

Required Qualifications

• BS, MS, or PhD in Robotics, Computer Science, Machine Learning, Electrical Engineering, Mechanical Engineering, Aerospace Engineering, or a related field.
• 4 years of hands-on experience building simulation environments and training/deploying robot policies on real hardware (an MS/PhD with equivalent research and project experience: roughly BS 4 years, MS 2 years, or a PhD in a relevant area).
• Deep experience with at least one major robotics simulator (Isaac Sim/Lab, MuJoCo, Drake, Gazebo, SAPIEN, Genesis, or similar), not just using it, but building and tuning environments in it.
• Demonstrated sim-to-real transfer: you've taken a policy trained in sim and made it work on a physical robot, and you understand why the gap exists and how to close it (domain randomization, system ID, calibration).
• Strong grasp of modern robot learning: reinforcement learning and/or imitation/diffusion policies, with working familiarity with VLAs and/or world models.
• Experience scaling data generation and collection for policy training (synthetic data, procedural generation, teleop/demonstration pipelines).
• Proficient in Python (and comfortable dropping into C where it counts); fluent with PyTorch or JAX and the modern ML tooling stack.
• A track record of getting robots actually to do the task in the real world: deformable or rigid/articulated manipulation on arms, bimanual setups, or humanoids.

Nice to Have (Strong Pluses)

• Using agentic / LLM-based tooling to automate sim and data work: reward generation (Eureka-style), automated scene/task generation, or agentic evaluation and curriculum design.
• Deformable object manipulation and the sim-to-real challenges that come with it (cloth, cable, soft bodies, contact-rich tasks).
• Experience with VLAs (e.g., OpenVLA, π0, RT-style policies) or world models for manipulation.
• Bimanual or whole-body / wheeled-humanoid manipulation.
• GPU-accelerated / massively parallel simulation and large-scale RL training.
• Real-time / on-robot deployment experience and familiarity with the controls and hardware stack, so your policies play nicely with the low-level controllers.
• Publications, open-source contributions, or competition results in robot learning or simulation.