Research Scientist: Pretraining

Sourced from jobs.ashbyhq.com

Generalist is building embodied foundation models to bring general intelligence to the physical world, starting with robotic dexterity. As a Research Scientist focused on pretraining, you will design and run large-scale training of robot foundation models across multimodal data spanning video, proprioception, action, language, and more. You will define architectures, objectives, and data strategies that give models new general capabilities across embodiments and tasks. The role sits at the intersection of frontier AI research and robotics, working closely with ML infrastructure and systems teams.

• Design and execute large-scale pretraining runs for robot foundation models using transformer- and diffusion-based architectures
• Define model architectures, objectives, and training curricula across multimodal robotic data (vision, action, state, language)
• Develop scalable data mixtures and sampling strategies across petabyte-scale datasets
• Guide data collection operations and source new datasets
• Run ablations to understand scaling laws, data quality effects, and architecture tradeoffs
• Collaborate with ML Infra and Systems to improve cluster utilization, throughput, and reliability
• Turn raw robotic interaction data into generalizable model capabilities

• Senior/staff (L6+) experience working on large-scale ML systems
• Deep experience training large transformer or diffusion models at scale (e.g. language, audio, or video models)
• Led or significantly contributed to multi-node, multi-GPU distributed training efforts
• Experience with scaling laws, optimization dynamics, and large-model failure modes
• Strong PyTorch fundamentals and comfort debugging across the full stack
• Strong empirical rigor and fast iteration speed

• Experience with multimodal or embodied foundation models
• Background in robotics or vision-language-action models
• Prior work at a frontier AI lab