Machine Learning Engineer - Sim2Real & Machine Modeling

Autonomy team at Gravis heavily relies on simulation to develop autonomous controllers. Whether these controllers work on the machine depends on how well we close the sim2real gap. In this role you will help us bridge the gap. We are looking for someone with strong ML/RL background and experience with real robotic systems.

Responsibilities

Machine & dynamics modeling

• Build ML models to help bridge the sim2real gap
• Decide what architecture the problem actually needs - sequence models, state-space formulations, something else - and back the choice with data
• Characterize where the gap actually comes from: which unmodeled effects hamper the sim2real transfer, and which we can safely ignore
• Answer how much data is needed and what distribution it has to cover

Performance monitoring

• Define the performance metrics and validation methodology for model fidelity and sim2real transfer
• Build models and methods that detect machine properties changing over time
• Work closely with the autonomy and simulation teams — your models influence the controllers that run on the machine

Required Qualifications

• Degree in Computer Science, Robotics, Machine Learning, Engineering, or a related field
• Strong Python and PyTorch, strong git skills
• Solid experience modeling time-series or dynamical-system data from large datasets - sequence models, system identification, or state-space approaches
• Strong analytical skills: you design the experiment, run the ablation, and draw a conclusion you'd defend

Nice to have

• Reinforcement learning experience
• Imitation learning or learning from demonstration, especially from human operator data
• Familiarity with recent literature and methods in learned behavior policies
• Classical system identification, control, or hydraulics background
• You like to solve problems outside of the laboratory
• You like a culture where the best idea wins no matter whether it comes from the CTO or an intern, as long as it's backed by numbers
• You are comfortable owning the result end to end: when the data you need doesn't exist yet, you go on site, touch the machine and get it
• You'd take a simple model that measurably closes the gap over a sophisticated one that might, and you're patient enough to get there in steps