Senior ML Engineer — Robotic Manipulation

Kenji Nakamura

Full-time · Senior · New_York

This role is no longer accepting applications

Similar open roles you can apply to now:

About the role

We are building robots that can handle the physical tasks in warehouse fulfilment that have resisted automation for two decades: picking irregular items, handling soft goods, and placing objects precisely into mixed-SKU cartons. We have fifteen robots deployed in three live warehouse environments. They handle over 40,000 picks per day between them. The manipulation ML — grasp planning, object pose estimation, failure detection, and adaptive re-grasp — runs on our robots in real time with a 200ms end-to-end budget. We are not a university lab. We deploy into operational warehouses where downtime has real costs and robot errors have physical consequences. We are looking for a Senior ML Engineer to own the perception and manipulation learning pipeline — from training data collection through to model deployment on the robot edge. The work requires someone who takes both the ML and the systems engineering seriously, because a model that is accurate in simulation and fragile on the factory floor is worthless to us.

Responsibilities

Own the manipulation ML pipeline from synthetic and real training data generation through to on-robot deployment
Improve grasp success rate on irregular and deformable items — the hardest subset of our SKU range
Build failure detection and re-grasp decision models that recover gracefully when the primary grasp attempt fails
Develop evaluation methodology that bridges the gap between offline model metrics and on-robot pick success rate
Work with robotics engineers and warehouse operations teams to understand failure modes and translate them into training data and model improvements

Requirements

5+ years of ML engineering with at least 2 years building computer vision or manipulation models deployed on physical robotic systems or safety-critical edge hardware
PyTorch for developing and fine-tuning perception models for 3D scene understanding, object pose estimation, and grasp quality prediction
Python for training data pipelines, model evaluation, and edge deployment tooling — you write production-grade code that a mechatronics engineer can also read
Computer Vision at depth: RGB-D image processing, point cloud analysis, and the practical challenges of domain shift between controlled training environments and live warehouse conditions
Docker for reproducible training environments and containerised edge deployment on robot compute hardware
ROS or equivalent robotics middleware experience is a significant advantage
Statistics for rigorous evaluation: the difference between simulation accuracy and real-world pick success rate, and why they diverge

Benefits

$148,000 – $185,000 base salary + equity (Series B, strong traction)
Hybrid — two days at our Boston robotics facility, three days remote — relocation support available
$2,500 annual research and equipment budget
Direct feedback loop between model changes and real-world robot performance — measured in picks per hour, not validation loss
25 days PTO plus company holidays