Research Scientist, VLA Models - Atlas

Boston Dynamics•Waltham, MA

1d•$177,000 - $225,000

About The Position

As a VLA Research Scientist on the Atlas team, you will architect, train, and deploy large-scale behavior models that enable high-performance dexterous manipulation on Atlas. The work focuses on building models that take multimodal input—including vision, language, tactile sensing, task context, and robot state—and produce grounded actions that generalize across manipulation tasks, embodiments, and environments. You will design large-scale behavior cloning and imitation learning pipelines and integrate learned components into a complex, real-world robotic system. You will collaborate closely with robotics, controls, software, and hardware teams, and rapidly test your work on a state-of-the-art humanoid robot.

Requirements

MS with 3+ years of experience, or PhD, in Machine Learning, Robotics, Computer Science, or a related field.
Prior experience training and deploying learned policies for complex behaviors in robots or simulated characters.
Strong background in behavior cloning or imitation learning.
Strong background in diffusion policies, World Actions Models, and other state-of-the-art methods.
Strong background in multimodal robot learning involving vision, language, action, and tactile sensing.
Experience with modern ML frameworks such as PyTorch or JAX, simulators, and large-scale training workflows.
Strong analytical and debugging skills.
Ability to write and review well-structured research code.

Nice To Haves

Hands-on experience deploying learned policies on real robots.
Experience creating or curating large robotics datasets, including teleoperation, UMI, or egocentric video.
Experience with reinforcement learning as a complementary tool within behavior learning pipelines.
Publications in top-tier ML or robotics conferences such as CoRL, RSS, ICRA, or NeurIPS.
Experience contributing to large C++ or Python codebases.

Responsibilities

Architect and train end-to-end VLA and Large Behavior Models for mobile manipulation on Atlas.
Build large-scale imitation learning pipelines using human demonstrations, teleoperation, and simulation data.
Work on large-scale pre-training across multiple data sources and fine-tuning for precision tasks.
Develop policies capable of few-shot generalization across diverse factory manipulation tasks.
Integrate models into Atlas’s autonomy stack in collaboration with controls and platform teams.
Deploy and iterate on models using physical hardware and rigorous evaluations.
Write high-quality, maintainable Python code that fits into a large production codebase.