Applied Researcher - Deployment Intelligence

The role

You will own the gap between lab performance and customer site performance by mining fleet sensor and video data for failure modes. You will build monitoring systems and turn deployment data into continuous model improvements.

What you will do

• Design and ship pipelines that turn real deployment data into targeted fine-tuning and online policy improvement.
• Mine high-frequency multimodal sensor and video data to catch failure modes, drift, and regressions.
• Apply reinforcement learning to improve policies directly from real-world deployment data.
• Build automated monitoring that flags anomalies, near-failures, and out-of-distribution scenes in real time.
• Characterize and close generalization gaps as robots move to new sites.
• Build evaluation harnesses and data-selection strategies.
• Partner with Research, Data, and Deployment teams to turn findings into shipped improvements.

Required qualifications

• Bachelor’s, Master’s, or PhD in CS, Robotics, Statistics, or a related field, or equivalent practical experience.
• Hands-on experience in at least two of: reinforcement learning, sensor-data modeling/anomaly detection, vision-language models, or continual/online learning.
• Experience building monitoring, evaluation, or data pipelines for a live ML system.
• Comfortable designing and reading production experiments (A/B tests, canary rollouts, staged fleet deployments).
• Strong Python and PyTorch (or JAX).
• Comfortable with large multimodal datasets and distributed compute (Slurm/GPU clusters).
• Able to turn a fleet-scale data investigation into a clear recommendation.

Nice to have

• Experience with robot fleets or other physically-deployed autonomous systems in the field.
• Experience building or fine-tuning perception or foundation models for automated monitoring, captioning, or anomaly detection.
• Background in statistical methods for detecting anomalies and drift applied to sensor or telemetry data.
• Experience with human-in-the-loop learning: reward modeling from operator corrections, active learning, or data curation from failure cases.

About the company

Dyna Robotics builds general-purpose robots powered by a proprietary embodied AI foundation model. Their team comes from Google DeepMind, Meta, and Cruise.