2027 Internship Behavior ML Engineer, Learned Manipulation Policies

Bedrock Robotics Inc • San Francisco, CA

38d

About The Position

Teaching a 40-ton excavator to move like an expert operator is a very different problem from teaching a car to stay in its lane. The motions are multimodal, there are many good ways to swing, dig, and dump and the consequences of getting them wrong are measured in cubic yards and bent steel. Bedrock's Behavior ML team builds the learned policies that decide what our machines actually do, and diffusion policies are a central bet: models that can represent the full distribution of expert behavior instead of averaging it into mush. As our Behavior ML intern, you'll help train those policies and build the evaluation that tells us whether they're genuinely better working alongside two hosts who sit on both the training and evaluation sides of the problem.

Requirements

• Currently pursuing a BS, MS, or PhD in computer science, robotics, machine learning, or a related field or bringing equivalent research or industry experience
• Strong Python and hands-on experience training models in PyTorch (or equivalent)
• Working understanding of generative modeling diffusion models, flow matching, VAEs, or similar and of imitation learning or behavior cloning
• Experience running and interpreting real training experiments: you know how to tell a real improvement from noise
• Clear communication, you can explain what you tried, what happened, and what you'd do next

Nice To Haves

• Published work or substantial project experience in diffusion, flow matching, robot learning, or imitation learning
• Experience training and building Vision Language Action (VLA) models
• Experience training Reinforcement Learning (RL) policies for robot manipulation
• Experience evaluating policies in simulation and reasoning about the sim-to-real gap
• Familiarity with large-scale training infrastructure and experiment tracking tooling
• Exposure to robotics, autonomous vehicles, or other physical-world control problems
• Interest in construction, earthwork, or heavy equipment - no prior experience required

Responsibilities

• Train and iterate on diffusion flow matching, and related behavior policies using real fleet demonstration data and simulated rollouts
• Run architecture, conditioning, and hyperparameter explorations, and turn the results into clear findings the team can build on
• Build and improve evaluation pipelines that score behavior models on task success, smoothness, safety margins, and operator-likeness
• Investigate failure modes - distribution shift, mode collapse, out-of-distribution scenes and propose fixes
• Work with the simulation and eval teams to make sure offline metrics actually predict on-machine performance
• Contribute to the shared training codebase with clean, reviewed, reproducible work
• Share results regularly with the behavior, controls, and autonomy teams

Benefits

• We're committed to building a diverse and inclusive workplace.
• If you need an accommodation to participate in the application or interview process, please let your recruiter know so we can support you.

What This Job Offers

• Career Level: Intern
• Education Level: No Education Listed