Robot Learning Engineer

Lightspeed•Northbrook, IL

2d•$100,000 - $120,000

About The Position

As a Robot Learning Engineer, you will teach our robots to perform complex manipulation tasks from human demonstration. You will own the path from demonstration to deployed behavior: building the teleoperation systems that capture how skilled people do the work, turning that data into trained policies, and proving those policies on real robots with dexterous hands and custom tooling. You'll also shape how we learn from video of people doing the work itself. This role is hands-on, fast-moving, and measured by one thing: robots that reliably do the job.

Requirements

3+ years of hands-on experience in robot learning, ML for robotics, or closely related production ML
Experience training policies from demonstrations on physical robots using imitation learning (e.g., ACT, Diffusion Policy, or similar)
Experience collecting teleoperation or demonstration data for robot learning
Strong PyTorch skills and experience running training, evaluation, and debugging end to end
Track record of evaluating models in the real world and diagnosing failures
Strong Python; working C++
Experience with ROS or ROS2 and real robot hardware
Understanding of manipulator kinematics and control fundamentals
Proficiency with Linux, Docker, and Git

Nice To Haves

MS or PhD in Robotics, Machine Learning, Computer Science, or a related field
Experience fine-tuning VLA or foundation models for robot control
Experience with dexterous hands, VR hand tracking, or motion retargeting
Experience building or using leader-follower teleoperation systems
Experience learning from human video or egocentric data
Experience building and maintaining large-scale robotics datasets
Robotics simulation experience with MuJoCo, Isaac Sim, or similar
Experience with robot learning datasets and tooling (e.g., LeRobot, rosbag, MCAP)
Experience managing data vendors or labeling workflows
Background in manufacturing, industrial automation, or construction technology

Responsibilities

Build and improve teleoperation pipelines using leader arms for bulk arm motion and VR hand tracking for fine, finger-level interaction
Retarget human hand motion to a dexterous robot hand, developing in simulation first and validating on hardware
Design data collection protocols so demonstrations are consistent, well-structured, and train well
Define what data is worth collecting and how much, and guide technicians who run collection sessions
Train and fine-tune manipulation policies using imitation learning and vision-language-action (VLA) models
Choose policy architectures and action representations suited to multi-step, contact-rich tasks
Combine teleoperation data with human demonstration video to improve policies
Iterate quickly from data to trained policy to real-robot trial
Define and run evaluation protocols on real robots, measuring success rate per task step
Diagnose failures across data, model, and hardware, and fix the right one
Deploy policies on robots with ROS2-based control and optimize inference for real-time performance
Work with integration engineers to run policies across a full task sequence
Define dataset formats and data quality standards for training
Supervise external partners who build data pipeline infrastructure and label human demonstration video, and set specs and accept deliveries
Prototype extraction of useful training signals from egocentric (head-mounted camera) video in-house before scaling it out