Robotics Intern, World Action Models for Robot Policy

The Mission

We are building general-purpose robots that can perform a wide variety of dexterous, multi-step tasks in everyday environments. Our Vision-Language-Action Models (VLAs) are foundation models that generate robot actions from sensor data and natural-language requests. They are trained on a large curriculum of robot demonstrations, combined with internet-scale text, image, and video data and procedurally generated simulation data.

A growing focus of the team is world action models: models that learn to predict both how the world will change and which actions the robot should take. Because they capture the consequences of actions, these models can support planning over long horizons, recovery from errors, and anticipation of what nearby people are likely to do.

The Challenge

We envision robots that help with household chores and cooking, support older adults in living independently, and work side by side with people. These settings are messy and unstructured, and useful tasks often take minutes rather than seconds. A robot must chain many skills together, notice and recover from mistakes, and adapt to the people around it.

Our central question is what it will take for humanoids or other mobile manipulators to complete such long-horizon tasks reliably, safely, and naturally alongside people. We believe world action models, grounded in large-scale physical interaction data, are a key part of the answer for robot policy learning.

The Internship

We are seeking a research intern to adapt pre-trained world foundation models-through mid-training, fine-tuning, or post-training techniques-into world action models that serve as robot policies. These policies will be deployed on humanoid platforms or other mobile manipulators functioning within human-centered environments. Project focus areas can be customized based on your research interests, spanning model architecture design and training, long-horizon task execution, dynamic human-robot interaction, or real-world physical robot evaluation.

You will work closely with researchers and engineers on the team to turn ideas into working prototypes, run experiments in simulation and on real robots, and publish results at peer-reviewed venues. You should be comfortable working with large existing datasets as well as a growing corpus of robot and human-interaction data.

• Areas of Focus
• World action models for long-horizon tasks.
• Human-robot interaction and collaboration.
• Humanoid whole-body manipulation.
• Cross-embodiment transfer and learning from human video.
• Sim-to-real and synthetic data.
• Post-training and continual learning.

Responsibilities

• Conduct original research on world action models for robot policy, from problem definition through experiments and analysis.
• Mid-train, fine-tune, or post-train world action models, and evaluate them in simulation and on humanoids or mobile manipulators.
• Design evaluations that measure long-horizon task success and the quality of interaction with people.
• Collaborate with researchers and engineers across teams, the Robotics division, and our university partners.
• Share results through internal presentations, peer-reviewed publications, and open-source releases where appropriate.

Qualifications

• Currently pursuing a Ph.D. or M.S. in Robotics, Computer Science, Machine Learning, Mechanical Engineering, or a related field.
• Experience training or fine-tuning large models (e.g., VLMs, VLAs, or video/world models) on large datasets.
• Strong Python skills and fluency with a deep learning framework such as PyTorch. C++ experience is helpful but not required.
• A "make it happen" attitude and comfort with fast prototyping and informative experiments.

Bonus Qualifications

• Publications at top-tier robotics or ML venues (e.g., RSS, CoRL, ICRA, IROS, NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, HRI).
• Experience deploying learned policies on real robots, especially humanoids or mobile manipulators.
• Background in world models, video prediction, or any other model-based planning.
• Experience with human-robot interaction, human motion or intent prediction, or learning from egocentric human video.
• Experience with long-horizon or hierarchical task execution.
• System integration skills across ML tooling, simulation, robot middleware, and data pipelines.

Pay and benefits

The pay range for this position at commencement of employment is expected to be between $40 and $58/hour for Massachusetts-based roles. Base pay offered will depend on multiple individualized factors, including, but not limited to, a candidate's experience, skills, job-related knowledge, and market location. TRI offers a generous benefits package including medical, dental, and vision insurance, and paid time off benefits (including holiday pay and sick time). Additional details regarding these benefit plans will be provided if an employee receives an offer of employment.

Please reference this Candidate Privacy Notice to inform you of the categories of personal information that we collect from individuals who inquire about and/or apply to work for Toyota Research Institute, Inc. or its subsidiaries, including Toyota A.I. Ventures GP, L.P., and the purposes for which we use such personal information.

TRI is fueled by a diverse and inclusive community of people with unique backgrounds, education and life experiences. We are dedicated to fostering an innovative and collaborative environment by living the values that are an essential part of our culture.

It is unlawful in Massachusetts to require or administer a lie detector test as a condition of employment or continued employment. An employer who violates this law shall be subject to criminal penalties and civil liability.