Senior Robot Learning Engineer

Lightspeed•Northbrook, IL

About The Position

As Senior Robot Learning Engineer at LightSpeed Build Technologies, you will own how the company's robots learn and perform physical skills. You will build the data pipeline for capturing human demonstrations, the policies that learn from that data, and the executor that runs classical or learned actions on the robots through a shared interface. The robots perform contact-rich, high-variance construction tasks, including taping seams, routing cable, finishing drywall, and setting fixtures. The role's systems will determine how quickly new trades can be taught and how reliably they perform in deployment. You will define demonstration-capture approaches, own the robot-learning data pipeline, set policy architecture, and build training and evaluation loops in simulation and on real robots. You will also own the action executor and contact-compliance layer, develop recovery behaviors, and deploy policies within real-time and safety limits. Working with the world-sensing and agentic-planning principals, you will co-own foundation interfaces and own the data and evaluation interface. For each new trade, you will guide demonstration collection, policy training and evaluation, and recovery development. You will set robot-learning and control architecture across both robot families, lead and mentor learning and manipulation engineers, and make the system usable by engineers teaching new trades without your involvement. The role remains hands-on, including writing production code and testing on real robots every week.

Requirements

At least 5 years of experience in robotics or machine learning, including learned manipulation policies deployed on real hardware.
Deep experience with imitation learning and reinforcement learning, including when to use each.
Experience with force and impedance control for contact-rich manipulation.
Experience building a robot-learning data pipeline covering teleoperation capture, labelling, versioning, and evaluation.
Deep-learning experience in PyTorch, from dataset development through deployed policy.
Strong Python and C++ skills.
Experience with ROS2, Linux, Docker, and CI/CD workflows.
A record of setting technical direction across teams and mentoring engineers.

Nice To Haves

MS or PhD in Robotics, Machine Learning, Computer Science, or a related field.
Experience with vision-language-action models or other robot foundation models.
Experience with diffusion policies, ACT, or similar visuomotor policy architectures.
Experience manipulating deformable objects such as cable, textile, or tape.
Bimanual or dexterous manipulation experience.
Sim-to-real transfer experience with Isaac Sim, MuJoCo, or similar tools.
Experience with teleoperation systems such as leader-follower arms, VR, or egocentric capture.
Experience with robot-learning datasets in MCAP or rosbag.
Experience with contact-rich assembly, finishing, or construction tasks.
Background in construction, manufacturing, or industrial automation.

Responsibilities

Define demonstration capture using teleoperation rigs such as leader arms and VR, and egocentric capture of people performing tasks.
Own the data pipeline, including synchronized observations and actions, episode labelling, dataset versioning, and quality checks.
Determine data needs for each new task and establish how failures are labelled and fed back into the process.
Partner with the world-sensing principal on observation spaces and sensor streams.
Set policy architecture, including a shared vision and force backbone with task-conditioned heads.
Train policies using imitation learning, diffusion or transformer policies, and vision-language-action models; use reinforcement learning where rewards can be clearly defined.
Build training and evaluation loops in simulation and on real robots, using held-out evaluations to gate deployment.
Develop manipulation capabilities for deformable materials such as cable, tape, and membrane.
Set coverage and first-pass-yield targets for each skill and measure performance against them.
Own the actuate executor, providing a common contract for classical control and learned policies so the planner can select either for each step.
Build the compliance layer for contact-rich work such as nailing, pressing, taping, mudding, and insertion, using force and impedance control.
Build recovery behaviors for common failures of each primitive.
Deploy policies on robots within real-time and safety limits.
Define what each tool must sense and report so its actions can be learned from and controlled.
Co-own foundation interfaces with the world-sensing and agentic-planning principals, and own the data and evaluation interface.
For each new trade, collect demonstrations, train and evaluate the policy, and add recovery behaviors.
Drive down the time required to teach a new trade.
Set robot-learning and control architecture across both robot families.
Lead and mentor learning and manipulation engineers.
Make the system usable by engineers to teach new trades without the senior engineer being in the loop.
Write production code and test on real robots every week.