Robotics Manipulation Engineer
Pubblicato il 10-09-2026 - Embodied Ai in Roma
About The Role We are looking for an engineer to own the path from data collection to a working robot, together with the interface someone can use to drive it. You will integrate a robot learning stack — LeRobot, DAgger, RL, ACT/diffusion policies, VLAs — into our distributed system, and own customer-facing manipulation capabilities end to end: choosing the approach, making the robot actually do the task, and being there when it is demoed.
This role is for library and framework integrators rather than from-scratch builders. We value engineers who pick the right off-the-shelf piece and integrate it cleanly. The work spans ML pipelines on one side, real product UI on the other, and a distributed system in between. Show and tell is part of the job, and we hire people who naturally build toward something visible.
You will join a fast-growing robotics and AI team building solutions to real-world problems.
What you'll do Own manipulation capabilities end to end: define the task, design the data collection, train and evaluate the policy, get it running on the robot,
and demo it live to customers.
Integrate robot learning libraries (LeRobot, DAgger, RL, ACT/diffusion policies, VLAs, NVIDIA Isaac) into our training and deployment pipelines.
Debug real manipulation failures — grasp failures, calibration drift, distribution shift between collection and deployment — and decide whether the fix is more data, a different policy, or a change to the setup.
Wire ML artifacts and tooling (W&B;, HF Hub, GCS) into the systems that actually run on robots.
Build and improve the teleoperation and data-collection interfaces that make good demonstration data possible.
Use simulation when it speeds things up and real robots when it does not.
What you bring A track record of integrating complex libraries and frameworks into working systems, including an understanding of their failure modes rather than only their happy path.
Experience integrating ML libraries into real software systems (researc