Robot Learning Engineer — Manipulation Full-time · San Francisco, CA View role →

Defexrobotics; San Francisco (CA); On-site; USD 150,000 - 180,000; Full time

Job summary

Defexrobotics is pursuing research on teaching robots assembly tasks from demonstrations and measurements. We’ll build policies and experiments to determine where learning helps the prototype and how to handle a different connector.

You will work across data collection, training and robot deployment with the founders and controls engineer; the first task is connector insertion, judged by the tested joint, recovery behavior and human intervention required.

Qualifications

• Trained a policy and tested it on a physical robot with analysis of failures.
• Ability to implement training/evaluation code in Python using PyTorch or JAX.
• Experience with imitation or reinforcement learning and data collection impact.
• Understanding of robot observations, actions and timing; debugging capabilities.
• Design experiments to separate model improvements from hardware changes.

Responsibilities

• Build demonstration and intervention tools recording synchronized data.
• Develop imitation-learning policies and explore RL for observed failures.
• Deploy models on the prototype, considering calibration and latency.
• Version datasets and checkpoints; create tooling to replay experiments.
• Evaluate policies against conventional control on test parts; analyze failures.
• Adapt approach to a second connector; measure additional demonstrations and effort.

Skills

Policy training

Python

PyTorch/JAX

Imitation Learning

Reinforcement Learning

Data collection

Robot debugging

Experiment design

Tools

C++

ROS 2

Edge Inference

Job description

About the role

We’re investigating how robots can learn assembly tasks from demonstrations, contact measurements and the results of physical tests. You’ll develop the policies and experiments that tell us where learning improves the prototype and what it takes to handle a different connector.

You’ll work across data collection, training and robot deployment with the founders and controls engineer. The first task is connector insertion; success is judged by the tested joint, the recovery behavior and the amount of human intervention. We’ll agree the start date around the prototype’s build schedule.

Pay

$150,000–$180,000 USD per year + equity.

Responsibilities

• Build demonstration and intervention tools that record synchronized observations, actions and physical test results.
• Develop imitation-learning policies and investigate reinforcement learning where it can address observed failures.
• Deploy models on the prototype, accounting for calibration, inference latency and the robot’s control limits.
• Version datasets and checkpoints, and build tools to replay attempts and compare experiments.
• Evaluate policies against conventional control on separate test parts, including failed insertions, damage and manual recovery.
• Adapt the approach to a second connector and measure the additional demonstrations, tooling changes and engineering time required.

Experience

• You’ve trained a policy, run it on a physical robot and investigated cases where it failed.
• You can build training and evaluation code in Python using PyTorch or JAX.
• Practical experience with imitation learning or reinforcement learning, including how data collection affects the result.
• An understanding of robot observations, actions and timing, and the ability to debug the system around the model.
• Experience designing comparisons that separate a model improvement from a change in parts, fixtures or test conditions.

Useful experience

• Assembly or insertion tasks, action-chunking policies or learning from human corrections.
• Force or tactile sensing, LeRobot or SERL.
• C++, ROS 2, edge inference or adapting policies across parts and fixtures.

First project

Establish a demonstration dataset and repeatable evaluation for the first connector. Train an initial policy, compare it with the programmed controller and use the recorded failures to choose the next experiment.