Senior Physical AI & Robotics Scientist - Lab Automation

Roche•South San Francisco, CA

1d•$168,100 - $312,300•Onsite

About The Position

Roche’s Research and Early Development organisations at Genentech (gRED) and Pharma (pRED) are leveraging advances in AI, data, and computational sciences to transform drug discovery and development. The new Computational Sciences Center of Excellence (CoE) is a strategic, unified group focused on harnessing the power of data and Artificial Intelligence (AI) to assist scientists in delivering innovative medicines. This role is part of the initiative to apply Physical AI to build fully autonomous, self-healing laboratories for lights-out drug discovery. The Computational Sciences CoE and the Data and Digital Catalyst (DDC) team are building the foundational platform for autonomous labs, translating AI into physical action to accelerate the delivery of life-saving medicines. As a Senior Physical AI & Robotics Scientist, you will be the technical lead and scientific architect for deploying Physical AI. This involves implementing 3D perception, multi-physics digital twins, edge vision, and closed-loop control on robotic hardware using technologies like NVIDIA Isaac, Omniverse, and Metropolis. The role offers an opportunity to pioneer the next generation of autonomous laboratories by bridging foundation models with real-world hardware, enabling robots to gain spatial awareness and resilience to autonomously solve problems in a lab environment.

Requirements

MS in Computer Science, Robotics, Electrical Engineering, Mechanical Engineering, or related technical field with 5+ years of experience, or a PhD with 2+ years of experience in deploying Physical AI or robotics solutions.
Strong interest in translating physical AI capabilities into biological workflows and drug discovery pipelines.
Proven track record of taking real-world robotic manipulation and sim-to-real systems beyond demonstrations into sustained operation.
Strong background in 3D computer vision, kinematics, motion planning (ROS 2, MoveIt 2, cuMotion), and modern machine learning frameworks (PyTorch) with strong C++/Python skills.
Hands-on experience with robotics simulation and robot-learning frameworks (NVIDIA Isaac Sim/Lab, Newton, MuJoCo, Drake, Gazebo) and sim-to-real validation methods.
Experience training or fine-tuning vision/multimodal models (VLMs, Cosmos) and optimizing neural networks with TensorRT for edge hardware (Jetson Thor/Orin).
Proficiency with visual servoing, closed-loop controllers, and hardware API integration (gRPC, SiLA 2, vendor SDKs).

Nice To Haves

Hands-on experience with OpenUSD scene graphs, custom USD schemas, or NVIDIA Omniverse workflows.
Experience working with high-throughput lab automation hardware, manipulators, or liquid handlers (e.g., Hamilton, Tecan, Precise Automation, UR arms).
Background in imitation learning, reinforcement learning, or robot foundation models for contact-rich manipulation and recovery.
A public portfolio of real-world projects on GitHub/GitLab showcasing work in robotics, simulation, or physical AI.

Responsibilities

Develop 3D computer vision perception pipelines (point clouds, depth maps, pose estimation) and dynamic motion planning algorithms (ROS 2, MoveIt 2, cuMotion) for reliable arm manipulation, hand-eye calibration, grasping, and AMR navigation in dynamic workcells.
Build multi-physics simulation environments (NVIDIA Isaac Sim/Lab, Newton Engine) to model robot-labware interactions, evaluate LLM-generated execution plans within digital twins, and train policy recovery models.
Apply system identification, domain randomization, and sim-to-real tuning to ensure simulation-trained control policies execute predictably on physical hardware.
Fine-tune multimodal Vision AI models (VLMs, NVIDIA Cosmos) using synthetic failure datasets (Isaac Sim) for real-time detection of lab anomaly modes (spills, bubbles, dropped microplates, tilted lids, tip-pick misses).
Optimize low-latency inference with TensorRT on NVIDIA Jetson (Thor/Orin) edge hardware, combining visual servoing and closed-loop feedback control for self-correcting systems.
Harness OpenUSD scene graphs to represent workcell geometry and state, integrating synchronization services to link executable digital twins across heterogeneous hardware APIs (SiLA 2, gRPC, vendor SDKs).