Applied Scientist II - Robotics Simulation, Robotics R&D

Applied Scientist II – Robotics Simulation, Amazon Robotics R&D

We are looking for an Applied Scientist to join the Robotics Simulation team at Amazon Robotics. In this role you will design, build, and validate the simulation environments and policy training pipelines that enable robots to learn manipulation and mobility skills in simulation and transfer them to real hardware.

You will work at the intersection of robotics simulation science and modern Physical AI: building GPU‑accelerated RL environments, implementing imitation learning workflows, characterizing sim‑to‑real gaps, tuning physics parameters against real‑world data, and evaluating learned policies both in simulation and on physical robots. You will collaborate closely with SDEs who build platform infrastructure, Technical Artists who create simulation assets, and partner science teams who consume your environments and pipelines for their model development.

This is a hands‑on, execution‑focused role. You will own specific simulation science deliverables end‑to‑end, from environment design through policy evaluation, with increasing scope and independence over time. You will contribute to technical design discussions, propose improvements to the team’s simulation fidelity and training methodology, and help establish best practices for robot learning in simulation.

• Design and implement GPU‑accelerated reinforcement learning and imitation learning environments in NVIDIA Isaac Lab for manipulation and mobility tasks.
• Build and maintain policy training pipelines supporting diverse model architectures (diffusion policies, VLAs, behavior cloning, actor‑critic RL) and evaluate trained policies in simulation.
• Characterize and reduce sim‑to‑real gaps through systematic validation: compare simulated sensor outputs, kinematics, and dynamics against real‑world robot data, then implement targeted improvements.
• Implement domain randomization strategies (visual, physics, geometric) to improve policy robustness and transfer to real hardware.
• Develop sim‑to‑real transfer techniques including system identification, physics parameter calibration, and visual domain adaptation.
• Create robot embodiment validation tests (joint kinematics, actuator response, contact behavior) to ensure digital twins are faithful to real hardware.
• Build data pipelines for recording, replaying, and augmenting demonstration data (from teleoperation or automated trajectory generation) to scale training data volume.
• Contribute to end‑effector modeling and contact dynamics tuning, ensuring physically plausible gripper and tool interactions in simulation.
• Author design documents for new simulation science capabilities and contribute to technical reviews.
• Collaborate with partner science teams to understand their model architectures and ensure simulation environments meet their training requirements.

• Medical, Dental, and Vision coverage
• Maternity and Parental Leave Options
• Paid Time Off (PTO)
• 401(k) Plan

• PhD or Master’s degree
• Knowledge of ML frameworks including JAX, PyTorch, vLLM, SGLang, Dynamo, TorchXLA, and TensorRT
• Experience in robotics design, automation systems development, control systems design, or related product development
• 2+ years of experience working with physics simulation platforms for robot learning (Mu Jo Co , Isaac Sim/Lab, PyBullet, Drake, or equivalent)
• Demonstrated experience training robot policies using reinforcement learning or imitation learning and evaluating them in simulation
• Experience with articulated robot simulation, including URDF/MJCF/USD formats and rigid/soft body dynamics
• Familiarity with sim‑to‑real transfer concepts (domain randomization, system identification, or physics calibration)
• Hands‑on experience deploying learned policies on real robot hardware (manipulation arms, mobile platforms, or mobile manipulators)
• Experience with NVIDIA Isaac Lab/Sim, Omniverse, or USD‑based simulation workflows
• Experience with modern Physical AI architectures: vision‑language‑action models, diffusion‑based policy learning, action‑chunking transformers, or behavior cloning from demonstrations
• Familiarity with…