Physical AI Senior - Manipulation: Drive our autonomous manipulation effort end to end: policy stack, data, evaluation discipline and architecture

About the role

Robots that restock shelves autonomously, at a reliability level a customer signs off on — that's the mission of this role. You will lead our autonomous manipulation effort end to end: the policy stack, the data behind it, the evaluation discipline that keeps the numbers honest, and the architectural direction of the whole track. You'll work with experienced VLA engineers who are ready to go deeper with the right technical lead.

What you'll do

• Own the manipulation policy stack from data collection through training to on-robot evaluation

• Track frontier AI manipulation research, test it on our hardware, and decide what to adopt, fine-tune, or build

• Own the measured success rate — and the methodology that makes it trustworthy

• Train models that are robust to sensor noise, partial observability, contact dynamics and environment variability

• Collaborate with the rest of the tech stack teams to drive the development of our humanoid software

• Ensure grasp quality and contact behaviour from the manipulation side

• Mentor and technically lead the engineers on the manipulation track

What we look for

Some combination of the following:

• You have trained and deployed a successful manipulation policy on real hardware

• Strong real-robot evaluation discipline — you understand how and why benchmark performance differs from field performance, and how to close that gap

• A data-first debugging mindset: when a policy underperforms, you look at the data before the model

• The seniority to extend and improve an existing codebase, and to raise the technical bar around you

• Hands-on experience developing and deploying robot learning algorithms on hardware

• A genuine drive towards reliability in the field

Tech stack

• Python, PyTorch, VLA (vision-language-action) model families, RL

• On-robot inference deployment (ONNX/TensorRT-class)

• Data logging and real-robot evaluation pipelines

• Git, GitHub, Docker and CI for reproducible training and deployment

• Training infrastructure for imitation and policy learning at scale

• Expertise working in Linux

Bonus experience

• Humanoid or bimanual manipulation

• Retail or logistics industry

• Perception knowledge: object pose estimation, depth estimation models, success detection

• Grasping based on touch sensors

• Simulation tools like NVIDIA Isaac Sim or MuJoCo