Robotics Researcher, World-Action Models (first robotics research hire)
San Francisco, on-site 5 days a week. Up to around $300k base plus equity.
The company: A Series A multimodal data lab of about 30 people, backed by Tier 1 investors and already generating strong revenue from top AI labs. It curates video, egocentric, tactile and 3D training data, and is moving from generative AI into physical AI.
The role: Train and evaluate world-action models and 3D perception systems to show which data actually improves robot capabilities. You own projects from question to answer, report to the CTO, and set the research direction as the first hire in this area.
What you will do
- Train and post-train action-conditioned world models and policies (manipulation, commanded actions, camera motion).
- Run data ablations across egocentric human video, teleop and robot rollouts.
- Build 3D labelling pipelines from raw video: depth, camera trajectories, hand/body/object pose, point tracks, reconstruction.
- Design held-out evaluations and benchmarks; use simulation where it makes them cheaper.
- Turn findings into curation and capture specs, and publish papers, datasets and benchmarks.
What you need
- 4+ years in robot learning, embodied AI or 3D computer vision. PhD welcome, not required.
- Have trained a policy or world model (VLA, diffusion policy, imitation learning, action-conditioned video model) and tested it on unseen data.
- Hands-on 3D vision: multi-view geometry, depth, pose, tracking or reconstruction.
- Fluent Python with PyTorch or JAX; rigorous with baselines and ablations.
- Clear communicator.
Nice to have: learning from egocentric or human video and retargeting to robots; policies deployed on real hardware; MuJoCo or Isaac Sim; large-scale distributed training; early-stage startup experience.
