Research Scientist / Engineer: Large Behavior Models

← Back to Careers

Research Scientist / Engineer: Large Behavior Models

Seoul Gangnam (On-site)

• Full-time·Talent Pool / Year-round Interest

About The Role

We build the manipulation skills FRIDAY performs as a single model, learned from data at scale.

The team works along two axes. One is developing how we acquire and process training data at scale. The other is training the model that turns observations and instructions into robot action. That work spans data composition, model architecture, pre-training, post-training, action representation and tokenization, conditional generation, policy decoding, vision-language-action and world action models, world models, inference optimization, and evaluation on real robots.

The two axes close into a single loop. Where the model fails decides what we collect next, and the newly collected data updates the model. Designing and running that loop is a large part of the work.

The model has to run on a real robot, in an environment nobody controls. What we build is a skill a higher-level system can call, and that skill has to follow instructions, handle variation, recover from mistakes, and finish the task. Whether it meets that bar is how we judge the team's work.

Key Areas

Large-scale data collectionData flywheelLarge Behavior ModelVision-Language-ActionWorld Action ModelWorld modelsAction representationsMultimodal pre-trainingPost-trainingPolicy fine-tuningReal-robot evaluation

What You'll Do

• Design and operate the collection and cleaning pipelines that produce trainable data at scale.
• Take part in data collection directly - build the tools and protocols, run them in the field, and open the resulting data yourself to improve it.
• Train and improve large behavior models that turn observations and instructions into robot action.
• Design action representations, tokenization, temporal abstraction, conditional generation, and decoding, and choose between them through controlled experiments.
• Own post-training that adapts pre-trained models to real tasks: per-task fine-tuning, improvement from autonomous rollouts and failure data, RL-based post-training, and performance and stability tuning before deployment.
• Establish how the scale and composition of data affect performance, and use that result to decide what to collect next.
• Build evaluation that measures success rate, generalization, recovery, and latency on real robots, and feed the results back into data and training.
• Deploy and debug learned policies on FRIDAY under real-time constraints.

Required Qualifications

• 3+ years of experience in machine learning, robotics, or intelligent systems (5+ for senior), or equivalent results.
• Experience pre-training and post-training large models in PyTorch, JAX, or an equivalent framework.
• Experience running a learning system end to end on real robots or real-world data.
• Strong Python and software engineering skills, and the ability to close the gap between research code and deployed code.
• The ability to move between data, model, and evaluation, and carry a problem from formulation through to demonstration.
• A willingness to take part in data collection and processing directly, not only model training.

Preferred Qualifications

• Experience with VLA, diffusion/flow policies, world models, or cross-embodiment learning.
• Experience operating and scaling large data pipelines in production.
• Experience with RL fine-tuning, preference-based learning, or post-training of large models.
• Experience with distributed training, efficient fine-tuning, model compression, or real-time inference optimization.
• Experience deploying and debugging learning-based systems on real robots.
• Strong publications or open-source contributions in foundation models, multimodal learning, or robot learning.

What We Offer

• Competitive compensation based on experience, level, and technical impact.
• The tools, compute, robots, and equipment needed to do your best work.
• The opportunity to build, test, and deploy directly on FRIDAY and the Holiday Robotics full stack.
• Close collaboration with researchers and engineers across hardware, control, simulation, learning, perception, data, and operations.

This is an evergreen talent-pool posting. We review profiles continuously, and the timing and level of a formal hiring process may depend on team priorities and available openings. If your work matches our direction, we would still like to hear from you. You are welcome to apply even if you do not meet every qualification.

Apply