ML Engineer, Foundation Models

Humble Robotics•San Francisco, CA

About The Position

We’re looking for an ML engineer to design, train, and ship the vision-language-action (VLA) foundation model at the core of Humble’s autonomous driving stack. You’ll work across the full arc—from architecture decisions and large-scale training to closed-loop evaluation in simulation and deployment on our trucks. This is a rare chance to build a production VLA for autonomous freight from the ground up, with the freedom and responsibility that comes with a small team tackling a massive problem.

Requirements

MS or PhD in Computer Science, Machine Learning, Robotics, or a related field—or equivalent industry experience
Strong proficiency in PyTorch, distributed training, and GPU-accelerated workflows
Solid foundation in transformer architectures, attention mechanisms, and modern generative modeling (diffusion, flow matching)
Eligible to work in the United States

Nice To Haves

Experience building or contributing to end-to-end autonomous driving systems
Track record of publications at top ML/robotics venues (NeurIPS, ICLR, ICRA, CoRL) or significant open-source contributions
Familiarity with sim-to-real transfer, photorealistic simulation, or neural rendering for driving scenes
Experience with reinforcement learning, imitation learning, or learning from demonstration in embodied settings
Comfort operating as an early team member—high ownership, low ego, fast iteration

Responsibilities

Design and iterate on our VLA model architecture—including the VLM backbone, action decoder, and multimodal fusion pipeline
Build and optimize large-scale training infrastructure (distributed training, data pipelines, mixed-precision, efficient fine-tuning)
Develop simulation-based evaluation and closed-loop training workflows using photorealistic neural rendering
Curate and manage multimodal training datasets spanning real-world driving and synthetic scenarios
Translate state-of-the-art research (diffusion/flow-matching action heads, reasoning-augmented VLAs, world models) into production-grade systems
Collaborate directly with vehicle systems and controls engineers to integrate model outputs into a real-time autonomous driving stack