Research Engineer

Pantheon is building general-purpose robots. Research velocity — experiments shipped per week — is the primary bottleneck on how fast Pantheon gets there. Research Engineers directly determine that rate.

The Opportunity

Pantheon is looking for a Research Engineer to work alongside researchers on large pretraining runs, RL post-training, and the data and infrastructure that make them scale. You will own the loop from the H100 training cluster to robots running on a factory floor.

This is a role for a strong software engineer first, with deep ML intuition — someone who has implemented models end-to-end, not just run existing ones. If you want your engineering work to directly determine how fast a robotics frontier gets pushed, this is that role.

What You'll Do

• •      Implement model architectures and training recipes alongside researchers and make them run efficiently at scale


• •      Build and maintain data pipelines including collection, curation, filtering, and augmentation across vision, proprioception, action, and language


• •      Own training infrastructure including distributed training, checkpointing, profiling, and debugging across the GPU fleet


• •      Build evaluation that catches regressions and produces real signal, both offline and on-robot


• •      Own the loop from training cluster to deployed robot and close it with data from the field

You Should Have

• •      Strong software engineering foundations with deep ML intuition


• •      Experience implementing ML models end-to-end, not just running existing ones


• •      Fluency in PyTorch or JAX with hands-on distributed training experience


• •      Ability to debug across the full stack


• •      Comfort moving fast in ambiguity

Nice to Have

• •      Experience with large-scale training infrastructure including multi-node, multi-GPU, and cluster environments


• •      Prior work on large multimodal models


• •      Publications at NeurIPS, ICML, ICLR, CoRL, RSS, ICRA, or similar venues


• •      Experience deploying models on physical hardware and optimizing for latency and compute at the edge