Member of Technical Staff, Robotics Research

Sieve•San Francisco, CA

1d•Onsite

About The Position

Sieve is hiring its first Robotics Researcher to lead the training and evaluation of world-action models (WAMs) and 3D perception systems, investigating how data affects robot capabilities. The role connects the capture, curation, and labeling of physical-world data to measurable improvements in robot learning. The researcher will lead projects from identifying model limitations through selecting data, designing experiments, and evaluating results on tasks and environments not seen by the model. The work emphasizes converting raw multimodal and egocentric video into data grounded in geometry, pose, and action. Simulation is used when it makes evaluation or augmentation more efficient, rather than being the central focus. The role is based onsite at Sieve’s San Francisco headquarters, five days per week. Findings may be shared through papers, technical reports, blog posts, and, where appropriate, open datasets or benchmarks.

Requirements

At least 4 years of research or engineering experience in robot learning, embodied AI, or 3D computer vision.
Fluency in Python and PyTorch or JAX, including the ability to write, debug, and modify training code independently.
Hands-on 3D vision experience in areas such as multi-view geometry, depth, pose, tracking, or reconstruction.
Experience training a policy or world model, such as a VLA, diffusion policy, imitation-learning system, or action-conditioned video model, and testing it on data not used in training.
Strong experimental rigor, including setting baselines, running ablations, and determining whether gains came from data, architecture, compute, or evaluation leakage.
Comfort working with large, messy datasets and diagnosing data-quality issues.
Strong communication skills and the ability to explain findings, tradeoffs, and uncertainty clearly.
Ability to work in person at Sieve’s San Francisco headquarters.

Nice To Haves

A PhD is welcome but not required; strong publications, open-source projects, or systems the candidate can explain are also valued.
Experience with robot learning from egocentric or human video, including retargeting human motion onto robot embodiments.
Experience deploying a trained policy on real robot hardware and measuring it in closed loop.
Experience using simulators such as MuJoCo or Isaac Sim for evaluation or synthetic data.
Experience with large-scale distributed training.
Experience as an early hire at a startup.

Responsibilities

Compare data sources for robot learning, including egocentric human video, teleoperation, and robot rollouts, and investigate how data quality, mixture, embodiment, and supervision affect policies and world models.
Train and post-train action-conditioned world models and policies for capabilities such as following commanded actions, manipulating objects, and responding to camera motion.
Build 3D labeling pipelines for depth, camera trajectories, hand, body, and object pose, point tracks, and scene reconstruction.
Design evaluations by fixing test tasks, environments, and embodiments before model comparisons, and report clear, verifiable conclusions about the effects of data.
Use simulators or learned world models for evaluation, augmentation, and sim-to-real checks.
Translate research findings into curation filters, annotation specifications, and capture protocols for Sieve’s pipelines and data delivered to partner labs.
Collaborate with Sieve engineers and researchers at partner labs to select problems and share findings.
Share research through papers, technical reports, blog posts, and, where appropriate, open datasets or benchmarks, and contribute to how the field understands the role of data in robot learning.