Research Scientist, RL for Dexterous Manipulation, Atlas

Boston Dynamics, Inc.; Waltham (MA); On-site; USD 175,000 - 220,000; Full time

Benefits offered by this job

Generous benefits package

Access to a world-class fleet of robots

Annual bonus structure

Job summary

Boston Dynamics, Inc. is looking for an RL Research Scientist in Waltham, Massachusetts, to lead research projects focused on visual sim-to-real transfer and improve manipulation skills. Ideal candidates hold a PhD in a related field or have significant experience with reinforcement learning and robotics.

The role involves developing algorithms, refining models, and training to enhance dexterity in robotic systems. The compensation ranges from $175,000 to $220,000 annually, complemented by a generous benefits package.

Qualifications

• PhD or MS with 3+ years in ML or Robotics.
• First-author publications in top AI venues.
• Experience with training policies for manipulation.
• Hands-on with VLA models or large behavior models.
• Proficiency in PyTorch and/or JAX.

Responsibilities

• Develop algorithms for visual sim-to-real transfer.
• Design post-training recipes for VLA models.
• Research reward modeling and offline-to-online RL.
• Train policies that generalize across tasks.

Skills

Reinforcement learning

Algorithm development

PyTorch

JAX

Robotics

Education

PhD in ML, Robotics, or a related field

MS with 3+ years of experience

Tools

VLA models

Diffusion policies

Job description

Position Overview

Are you passionate about using reinforcement learning to solve dexterous manipulation tasks? As an RL Research Scientist, you'll lead research projects on visual sim-to-real transfer and post-training of Vision‑Language‑Action (VLA) models. Your job is to turn unlabeled data from simulation or the real-world into robust real‑world manipulation skills. You should push the frontier of what bimanual and multi‑fingered systems can do in unstructured environments.

Responsibilities

• Develop novel algorithms for visual sim-to-real transfer with photorealistic rendering.
• Design post-training recipes that improve pre-trained VLA models on manipulation tasks.
• Research reward modeling, and offline-to-online RL for large multimodal policies.
• Close the sim-2-real gap through tactile sensing, vision, and system identification.
• Train policies that generalize across objects, scenes, and embodiments.

Required Qualifications

• PhD in ML, Robotics, or a related field, or an MS with 3+ years of experience.
• Track record of first-author publications at top venues (CoRL, RSS, ICLR, NeuRIPS).
• Demonstrated experience training policies for dexterous or contact-rich manipulation.
• Hands‑on experience with VLA models, diffusion policies, or large behavior models.
• Proficient in PyTorch and/or JAX, with experience training models at scale.
• Strong software fundamentals and the ability to ship research code that runs reliably.

Preferred Qualifications

• Deployed vision-based manipulation policies on physical robots.
• Deep knowledge of sim-to-real transfer techniques and photorealistic rendering.
• Built training pipelines that combine RL, imitation learning, and large-scale pretraining.
• Experience fine-tuning foundation models with RLHF, DPO, GRPO, or related methods.
• Familiarity with tactile sensing, multi-fingered hands, or bimanual coordination.

Benefits and Compensation

• Compensation: $175,000 to $220,000 annually (base pay depends on internal equity, job-related knowledge, skills, and experience).
• Generous benefits package including medical, dental, vision, 401(k), paid time off, and an annual bonus structure.
• Access to a world-class fleet of robots, simulation infrastructure, and compute.
• Direct impact on the policies powering our humanoid robots in the real world.
• The chance to define what’s possible for dexterous manipulation at scale.

Company

Boston Dynamics is building a commercial company that delivers cutting-edge technology and solutions to our customers from industrial applications to logistics and warehouse solutions.