Research Scientist Intern, Robot Foundation Models (PhD)

Meta•Redmond, WA

1d•$7,650 - $12,667

About The Position

Agios is a full-stack research organization within Reality Labs Research, spanning materials and manufacturing, sensors and actuators, simulation, and learned control. Its mission is to build systems that enhance human capabilities through intelligent machine collaboration. The Dexterous Manipulation pillar pairs a custom dexterous robotic hand with in-house wearable sensing, high-accuracy simulation, and glove-driven teleoperation to address the gap between the accelerated pace of digital research and development and the slower pace of physical work. The Research Scientist Intern will work at the intersection of robot learning, dexterous control, and human-to-robot transfer, with mentors in hardware, simulation, controls, and policy learning. The work focuses on learned models that turn human demonstration data into executable robot behavior, including retargeting, manipulation policies, and teleoperation. Internships run 12 to 24 weeks, with various start dates throughout the year.

Requirements

Currently has or is in the process of obtaining a Ph.D. in Computer Science, Artificial Intelligence, Robotics, or a relevant technical field.
Understands the state of the art in learned dexterous manipulation and control and can judge which open problems are worth advancing.
Has experience with Python and PyTorch or JAX, plus familiarity with ROS or comparable robotics middleware.
Must obtain work authorization in the country of employment at the time of hire and maintain ongoing work authorization during employment.

Nice To Haves

A proven track record of significant results, demonstrated through grants, fellowships, patents, or first-authored publications at leading workshops or conferences in robotics, machine learning, or computer vision, or similar venues.
Experience training modern control policies, including flow matching, diffusion, imitation learning, reinforcement learning, and large behavior models.
Familiarity with robotics modeling and analysis, including kinematics, dynamics, trajectory optimization, motion planning, perception, task planning, and control theory.
Hands-on experience with multi-fingered robotic hand manipulation and with parallel-jaw or other grippers mounted on a robot arm.
Experience with motion retargeting across human and robot embodiments.
Experience with sim-to-real and real-to-sim workflows, including domain randomization, system identification, scene reconstruction, and closing the transfer gap.
Experience collecting and learning from real robot hand demonstrations through teleoperation, kinesthetic guidance, or wearable capture; bimanual setups are a plus.
Experience with multimodal or self-supervised representation learning and cross-embodiment transfer.
Experience with vision-based input recognition systems, such as hand tracking, pose estimation, and object pose estimation.
Experience working with time-sequence sensor data, such as tactile and IMU streams.
Experience with hand force estimation and its applications in robot policy learning.
Experience designing data collection protocols and building high-quality, diverse datasets.
Experience bringing up and debugging prototype software and hardware systems, such as robot platforms, multi-camera tracking rigs, or wearable sensing systems.
Experience working and communicating cross-functionally in a team environment.
Intent to return to the degree program after completion of the internship or co-op.
Availability for a minimum 16 consecutive-week internship.

Responsibilities

Drive an independent research thread on learned control and motion transfer for dexterous hands, from problem formulation through evaluation on a real robot or in high-fidelity simulation.
Build training pipelines for modern policy classes, including flow matching and diffusion policies, imitation learning, reinforcement learning, large behavior models, and VLAs.
Develop representations that fuse multimodal human and robot data—including video, hand pose, tactile and contact force, and IMU—into a shared action space.
Construct data collection pipelines and benchmarks for dexterous manipulation, including held-out and out-of-distribution protocols that test generalization across objects, motions, and embodiments.
Collaborate across an interdisciplinary team of hardware, simulation, controls, and policy researchers, and communicate research plans, progress, and results.