Robotics Data Engineer

Turn raw footage into something a VLA model can actually train on.

About the role

You will own the pipeline between what our capture network produces and what a research team can load. That means segmentation and step-level action decomposition, annotation schemas agreed with buyers, and the quality tooling that catches problems before a customer does — IMU and video out of sync, coverage gaps, label drift across suppliers. The data arrives from several sources with different rigs and habits, so a large part of the job is making inconsistent input come out the other side consistent.

What you'll do

• Build and run the action-labeling pipeline — segmentation, step decomposition, QA tooling
• Define annotation schemas with buyers and hold them consistent across suppliers
• Automate quality checks: IMU/video sync, coverage gaps, label drift
• Turn recurring failure modes into checks that run before delivery, not after

What we're looking for

• You have shipped data pipelines that fed real ML training, not just prototypes
• Comfortable with video and time-series data at volume
• Nice to have: egocentric video, VLA, or robot-learning datasets