About the job
Do you want to create the greatest-possible worldwide impact in Robotics? Amazon has the world's most exciting treasure trove of robotics challenges. At Amazon Robotics we build high-performance, real-time robotic systems that can perceive, learn, and act intelligently alongside humans—at Amazon scale. Amazon Robotics invents and scales AI systems for robotics in fulfillment. Our mission is to enable robots to interact safely, efficiently, and fluently high density real-world fulfillment centers. Our AI solutions enable robots to learn from their own experiences, from each other, and from humans to build intelligence that feeds itself.
Responsibilities
- Architect, design, and implement 3D perception models - including encoder-decoder networks, query-based transformers, and generative architectures- for semantic occupancy prediction and scene completion on robotic platforms.
- Own the end-to-end model lifecycle: develop scalable training pipelines, optimize inference latency for ARM-based edge processors, and deploy production models that meet real-time performance targets.
- Design and scale pseudo-ground-truth data generation pipelines - both heuristic-based and learning-based (e.g., SAM3D, shape completion) to produce curated training samples using SageMaker infrastructure.
- Drive multi-view perception integration by fusing multiple view camera inputs for robust 3D reconstruction in partially observed and occluded bin environments.
- Influence the team's technical strategy and contribute to the long-term vision and roadmap for 3D perception in fulfillment robotics.
- Partner with cross-functional stakeholders across engineering, science, and operations teams to define requirements, iterate on system design, and deliver end-to-end solutions from research prototype to production deployment.
- Maintain high standards by participating in design and code reviews, designing for fault tolerance and operational excellence, and creating mechanisms for continuous improvement.
- Prototype and validate concepts through simulation, synthetic data evaluation, and live robotic workcell testing using 3D metrics (mIoU, IoU) and affordance-based evaluation frameworks.
- Mentor applied scientists and engineers, raise the technical bar, and foster a culture of scientific rigor and rapid experimentation.
Qualifications
Minimum
- 4+ years of building machine learning models for business application experience
- PhD, or Master's degree and 6+ years of applied research experience
- Experience programming in Java, C++, Python or related language
- Experience with neural deep learning methods and machine learning
- Demonstrated expertise in 3D computer vision and deep learning for robotics - spanning semantic scene completion, occupancy prediction, depth estimation, multi-view reconstruction, and real-time model deployment on edge hardware.
Preferred
- Publications in top-tier venues (CVPR, ICCV, ECCV, NeurIPS, 3DV, CoRL) in 3D scene understanding, shape completion, or occupancy prediction.
- Deep expertise in generative 3D models, vision transformers, and semantic scene completion architectures.
- Experience building large-scale pseudo-ground-truth or synthetic data pipelines (100K+ samples).
- Proficiency in real-time model optimization (ONNX/TensorRT) and deployment on edge hardware.
- Strong foundation in 3D geometry, multi-view reconstruction, and sensor fusion.
- Track record shipping ML models into production robotic systems with hard latency constraints.
- Effective communicator across science, engineering, and operations stakeholders in fast-paced environments.