Simulation Engineer – Embodied AI

Role overview

This role creates the simulation environments and digital twins used to train and evaluate embodied-AI policies. The goal is not visual reproduction alone, but simulation that generates useful training experience, reveals meaningful differences between models, and transfers to physical performance. The engineer will decide which aspects of robots, objects, environments, physics, sensing, and control need to be modeled for learning and evaluation.

Responsibilities

- Build programmable simulation environments for training and evaluating embodied-AI policies, including digital twins of real robots, workspaces, objects, and operating conditions. - Define tasks, success criteria, environment variations, edge cases, and procedural scenario-generation systems. - Generate large-scale synthetic data and trajectories, and develop domain-randomization strategies that improve policy robustness. - Model relevant physics, contacts, sensors, cameras, actuators, coordinate systems, and control interfaces. - Connect simulation environments to policy-training pipelines and build scalable systems for running large numbers of episodes. - Measure and reduce sim-to-real gaps, optimize simulation throughput, and provide tools that help ML engineers create experiments and modify tasks.

Requirements

- Strong software-engineering skills and experience building simulation, physics, game-engine, 3D, or other systems involving physical interaction. - Solid understanding of coordinate systems, rigid-body dynamics, geometry, contacts, and related physical concepts. - Ability to create clean, reusable, programmable environments rather than one-off visual demonstrations. - Experience with Python; familiarity with a systems language such as C++ is valuable. - Strong debugging ability across simulation, software, and physical systems, with an understanding of which real-world details matter for learned behavior. - Comfortable collaborating closely with machine-learning engineers and robotics researchers.

Nice to have

- Experience with MuJoCo, Isaac Sim, SAPIEN, Bullet, or comparable simulation platforms. - Background in embodied AI, robot-learning environments, manipulation, grasping, locomotion, or other physical interaction problems. - Experience with synthetic data generation, domain randomization, sim-to-real transfer, USD, URDF, meshes, CAD assets, or robot descriptions. - Experience scaling simulation workloads across GPUs or compute clusters.

What success looks like

Data generated in simulation improves policies, simulated evaluations predict meaningful differences between models, and experiments run efficiently enough to accelerate research. Policies and improvements developed in these environments increasingly remain effective when transferred to physical robots.