Head of Robot Learning

Anvil Robotics•San Francisco, CA

5d•Onsite

About The Position

Anvil is seeking its first Head of Robot Learning to take cutting-edge research papers and implement them on Anvil's hardware within weeks, shipping them as polished demos and guides. The role involves shaping the data collection process for Anvil's Physical AI platform, which includes robotics hardware and software designed for accessibility. Anvil has already shipped over 200 robots and generated significant revenue, with a focus on in-house manufacturing and supply chain integration. This position offers unique leverage over data collection, allowing the Head of Robot Learning to specify hardware modifications (cameras, mounts, sync schemes) to improve data quality. The role will leverage factory access for real-world industrial task data, moving beyond recycled public datasets. The goal is to reach a scale comparable to projects like LeRobot's shirt-folding, focusing on collecting high-quality demonstrations tailored to specific tasks. The company currently lacks a dedicated ML function, with model training handled by founders and controls engineers. The Head of Robot Learning will be the first hire for this function and will be responsible for establishing the end-to-end pipeline from data collection to policy deployment on Anvil's arms. The role emphasizes replicating the best public work rapidly, publishing results as open-source demos and guides, and ensuring a high level of polish on all shipped projects, including clean repositories, honest success rates, and comprehensive guides.

Requirements

Personally trained and deployed imitation-learning policies (ACT, Diffusion Policy, VLA fine-tunes) on real robot arms, not just cloud benchmarks.
Fluent in the layer under the model, understanding action chunking, temporal ensembling, inference latency, control-loop frequency, camera synchronization, timestamp alignment, and joint-space vs. cartesian command interfaces.
Ability to replicate papers in weeks, identifying transferable components and unreported issues.
Strong data instincts to identify issues in teleop demonstrations (inconsistent grasps, occlusions, timing skew) before training.
Proven ability to finish projects with a guide, video, and a README, and to measure and defend success rates.
Honest reporting of results, including n, evaluation protocols, and failure modes, especially in public.
Comfortable being the sole ML person in a fast, lean, founder-led company, setting own agenda and shipping on a weeks-not-quarters cadence.
Master's or Bachelor's in CS, robotics, or a related field.
Demonstrated trajectory of 2-4 fast-growing years or 1-2 years on a steep curve, acting as the research engineer making lab/team work run on hardware and closing hard problems.

Nice To Haves

LeRobot or similar open-source contributions.
DAgger / interactive imitation learning experience.
Public demos that achieved significant reach.
RL fine-tuning on real hardware.

Responsibilities

Own training pipelines end-to-end, including teleop and UMI data ingestion, dataset formats, quality triage, training jobs, and an evaluation harness with honest success-rate protocols.
Replicate high-leverage public robot learning work (e.g., folding-class manipulation, VLA fine-tunes, diffusion policies) on Anvil hardware within weeks and ship as public demo videos and reproducible guides.
Validate the UMI pipeline end-to-end, proving or fixing the path from the handheld data collector to a working policy on an OpenARM, and translating findings into hardware revisions.
Develop the data flywheel by implementing DAgger/human-in-the-loop correction workflows on the teleop stack to improve policies from intervention data.
Scale data collection beyond oneself by designing protocols for dedicated data-collection operators in Anvil's factory and partner facilities.
Ensure a high polish bar for all shipped projects, including demo videos, guides, and reproducible repositories.