About Origin

Origin is building physical AI for the built world. Our robots autonomously finish building interiors at production quality. OG-1 is deployed on live NYC commercial construction sites today. Backed by Accel.

Our system runs a Multi Agent Action Expert architecture: classical precision algorithms orchestrated alongside learned policies. The job is systematically expanding the learned components while keeping the system production-safe.

The Role

You own the full lifecycle of learned components on OG-1: from data collection and model training through edge deployment on Jetson AGX Orin. Every research project will have a deployment milestone. This is not a lab position.

What You Will Do

• Train and deploy VLA models for contact-rich manipulation using our imitation learning infrastructure.
• Build the data flywheel: teleoperation pipelines (GELLO, SpaceMouse, VR), DAgger-style online correction, demonstration curation.
• Research and prototype world models for surface state prediction, spray dynamics, and anomaly detection.
• Design offline evaluation metrics that predict real-world finishing quality before deployment.
• Optimize models for edge: TensorRT compilation, latency profiling, memory budgeting on dual Jetson AGX Orin.
• Design the interface where learned policies propose actions and deterministic safety layers enforce constraints.

Requirements

• BS/MS/PhD in CS, Robotics, ML, or equivalent experience shipping learned systems on physical robots.
• Strong Python and PyTorch; comfort modifying research codebases (you'll work directly with open-source VLA implementations).
• Experience in at least two of: imitation learning, RL, vision-language models, robot learning from demonstration, sim-to-real.
• Track record deploying ML on real hardware: not just training to convergence, but debugging why the policy fails on the actual robot.
• Working knowledge of ROS2 or equivalent robotics middleware.
• Experience working with Simulation Systems like Isaac Sim.
• GPU profiling and optimization (TensorRT, ONNX, CUDA); you understand why 200ms policy latency kills contact control.

Strong Plus

• Hands-on with VLA architectures (π0/π0.5, OpenVLA, RT-2, Octo) or foundation model fine-tuning for robotics.
• Teleoperation data collection and DAgger/HG-DAgger pipelines.
• World model architectures (DreamerV3, V-JEPA, latent dynamics models).
• Construction, manufacturing, or contact-rich industrial domains.
• Publications at CoRL, RSS, ICRA, NeurIPS: valued but equivalent shipped work counts.