Robotics ML Engineer

Simulation and Robot Learning | Noble Machines

Advance industrial autonomy by turning simulated experience into reliable real-world capability.

About Noble Machines

Noble Machines builds multipurpose robots to support human workers in the world's toughest jobs—turning dangerous work from a necessity into a choice. Our work demands reliability, robustness, and readiness for the unexpected—on time, every time. We're assembling a mission-driven team focused on delivering real impact in heavy industry, from construction and mining to energy. If you're driven to build rugged, reliable products that solve real-world problems, we'd love to talk.

The role

Your defining contribution will be solving the sim-to-real challenge from the simulation side: understanding why policies succeed in simulation but fail on hardware, then improving the environments, data, and training conditions that determine transfer. You will build simulation into a dependable tool for developing industrial autonomy.

You will own the learning pipeline end to end, from simulation and data generation through policy training, evaluation, deployment, and iterative improvement on physical robots. The role calls for strong judgment about which aspects of simulation need greater fidelity, which need broader variation, and how to demonstrate that either change improves real-world performance.

This is a hands-on ML engineering role with research depth. You will work closely with AI, controls, hardware, and robot operations teammates, owning the learning loop while partnering on the robot, control interfaces, and deployment infrastructure.

What you will do

• Lead simulation-side sim-to-real development. Use hardware evidence to identify consequential gaps in visual observations, physics, contact, sensing, and control execution; improve simulation fidelity, calibration, and randomization to address them.
• Build and maintain simulation environments and data-generation pipelines for robot learning. Design representative tasks, variations, and evaluation conditions, and ensure generated data is suitable for training transferable policies.
• Train and adapt vision-language-action (VLA) policies, world-action models (WAMs), and other visuomotor policies using simulated and real robot data. Build reproducible experiments and make informed choices about data quality, coverage, and training methods.
• Close the learning loop between simulation and hardware: analyze failures, collect corrective data, retrain, and validate improvements. Apply DAgger-style data aggregation and human-in-the-loop learning where appropriate.
• Use imitation learning and reinforcement learning to improve policy performance, including RL fine-tuning and learning from real-world experience where appropriate.