Research Engineer, Robot Learning

Relari; San Francisco (CA); On-site; USD 140,000 - 190,000; Full time

Job summary

Relari is a small research and engineering startup in San Francisco focused on teaching robots to learn dexterous skills from human biomechanics. This full-time on-site role involves building and evaluating robotics foundation models, adapting vision-language-action and world models, and deploying policies to real robots.

You will work directly with the founders, own problems end to end, and iterate on experiments that advance robot performance in real hardware environments.

Qualifications

• Strong foundations in machine learning, robot learning, or computer vision.
• Experience building and evaluating learning systems in Python using PyTorch, JAX, or similar tools.
• Ability to turn open-ended technical questions into working experiments with clear conclusions.
• Comfort debugging real data and physical systems rather than only clean benchmarks.
• Experience with VLA models, video models, world models, imitation learning, RL, or large-scale model training is useful.

Responsibilities

• Own pre-train, post-train, and evaluation pipelines of robotics foundation models from diverse multimodal data.
• Develop and adapt vision-language-action models, video models, and world models for dexterous manipulation.
• Investigate representations and training methods for transferring human skills across robot embodiments.
• Build reproducible large-scale training, simulation, and real-world evaluation loops that show progress.
• Deploy policies on physical robots and diagnose failures across data, perception, learning, control, and hardware.
• Improve practical experiment tooling, including data pipelines, teleoperation, retargeting, visualization, and robot runtime software.
• Read relevant research, reproduce ideas, and design focused experiments around results.

Skills

ML

Robot Learning

Computer Vision

Python

Tools

PyTorch

JAX

Job description

About The Role

You will work across the complete learning-to-deployment loop: building representations from human data, training models, evaluating them in simulation and on physical robots, and using failures to decide what to try next. Our research spans vision-language-action models, video models, world models, and policies that learn from biomechanical signals. We are looking for depth in machine learning or robot learning—not expertise in every part of the stack—and the willingness to follow an idea all the way to robot performance.

What You'll Do

• Own pre-train, post-train, and evaluation pipelines of robotics foundation models from diverse multimodal human and robot datasets.
• Develop and adapt vision-language-action models, video models, and world models for dexterous manipulation.
• Investigate representations and training methods for transferring human skills across robot embodiments.
• Build reproducible large-scale training, simulation, and real-world evaluation loops that make research progress measurable.
• Deploy policies on physical robots and diagnose failures across data, perception, learning, control, and hardware.
• Improve practical experiment tooling, including data pipelines, teleoperation, retargeting, visualization, and robot runtime software.
• Read relevant research, reproduce promising ideas, and design focused experiments around the results.

What We're Looking For

• Strong foundations in machine learning, robot learning, computer vision, or a closely related field.
• Experience building and evaluating learning systems in Python using PyTorch, JAX, or similar tools.
• Evidence that you can turn an open-ended technical question into a working experiment and a clear conclusion.
• Comfort debugging real data and physical systems rather than working only with clean benchmarks.
• Experience with VLA models, video models, world models, imitation learning, reinforcement learning, or large-scale model training is useful; we do not expect depth across every area.

Working at Relari

Relari is a small research and engineering startup developing new ways for robots to learn dexterous skills from human biomechanics. Our founders have AI research roots at MIT and NVIDIA, along with autonomous-vehicle and robotics deployment experience at Pony ai and Dexterity. We are backed by top investors including Y Combinator, General Catalyst, and Soma Capital. You will work directly with the founders, own problems end to end, and test your ideas on real robotic systems. This role is full-time and on-site in San Francisco.