Senior ML Engineer, Foundation Models

Chef Robotics•San Francisco, CA

2d•Onsite

About The Position

Chef Robotics is accelerating the deployment of intelligent machines in the physical world, starting with food production — the sector facing the largest labor shortage in the U.S., with 1.14M unfilled jobs today and 3.1M projected by 2030. These roles can't be offshored, making robotics essential to keeping production onshore and strengthening America's manufacturing base. Our AI-powered robots automate food prep and assembly in commercial kitchens and food manufacturing, and have already produced over 110 million meals in production — generating the world's largest proprietary dataset for deformable food manipulation. Backed by investors including Kleiner Perkins, Construct, Bloomberg Beta, and Promus Ventures, and built by a team from Cruise, Zoox, Google, Tesla, and Amazon Robotics, Chef is rapidly scaling with multiple multi-year contracts and a mission to put an intelligent robot in every commercial kitchen. About the Role The next leap in food robotics won't come from hand-tuned policies for individual ingredients — it will come from foundation models that generalize across thousands of food types, kitchen configurations, and manipulation scenarios out of the box. At Chef, we're building that model: the Food Foundation Model. As a Senior ML Engineer, Foundation Models, you will work at the frontier of large-scale robot learning: training and fine-tuning the Food Foundation Model, building the data infrastructure that feeds it, and deploying it onto physical robots in production kitchens. You'll bridge research and engineering — translating advances from the latest policy learning, generative modeling, and world model literature into systems that handle real food, with real end effectors, at real throughput. Your models won't just benchmark well; they'll serve millions of meals. We are a small, high-ownership team. We work onsite five days a week and move with startup urgency.

Requirements

MS or PhD in Machine Learning, Robotics, Computer Science, or a related field — or equivalent industry experience
5+ years of experience implementing and deploying ML models for real-world robotics applications
Hands-on experience with large-scale model training: pre-training, fine-tuning, and post-training alignment pipelines
Familiarity with modern policy and generative model architectures — diffusion models, transformers, behavior cloning, or large-scale multimodal models
Strong PyTorch skills and experience building reliable, production-quality training and evaluation infrastructure
Solid software engineering fundamentals in Python; able to write maintainable code across research and production codebases
Track record of taking models from research prototype to deployed system on physical hardware

Nice To Haves

Experience with world models or generative models for robot planning and prediction
Background in large-scale distributed training (multi-node GPU clusters, FSDP, DeepSpeed)
Familiarity with simulation environments (MuJoCo, Isaac Sim, Genesis) for synthetic data generation and domain randomization
Experience deploying models to edge hardware (ONNX, TensorRT, quantization, performance profiling)
Prior work with contact-rich manipulation, deformable object handling, or food robotics
Publications at top venues: CoRL, RSS, ICRA, NeurIPS, ICML, ICLR

Responsibilities

Define the architecture, training objectives, and learning approach for the Food Foundation Model — evaluating tradeoffs across generalization, sample efficiency, and deployment constraints
Investigate and evaluate the latest foundation model architectures — including VLAs, world models, JEPA-style joint embedding models, diffusion policies, and emerging approaches — and assess their applicability to Chef's manipulation and generalization challenges
Design pre-training, fine-tuning, and alignment pipelines that improve the model's ability to generalize across new food types, kitchen configurations, and end effector types with minimal retraining
Develop evaluation frameworks that measure real-world generalization and long-horizon reliability — not just offline benchmark accuracy
Collaborate with the data and platform teams on training data requirements, augmentation strategies, and model serving constraints
Stay current with the research frontier — reading and critically evaluating recent work from CoRL, RSS, NeurIPS, ICML, and ICLR and forming clear views on what's relevant to production manipulation