Software Engineer (Manipulation)
Posted Updated
Description
- The Software Engineer - Manipulation is a core contributor to our robot’s ability to interact with the world reliably and precisely
- This role is responsible for developing and deploying learning-based manipulation policies, with a focus on post-training methods that adapt and improve those policies on real tasks and physical hardware
- This role will bridge the gap between cutting-edge research and scalable, reliable production software
- Working closely with the simulation, data, hardware, and systems teams, this engineer will take manipulation policies from demonstration data through post-training to robust deployment across our fleet
- Implement and deploy state-of-the-art learning-based manipulation policies to achieve reliable, world-class performance on real-world tasks with physical hardware
- Adapt and improve policies for new tasks and hardware using post-training methods such as RL fine-tuning, policy distillation, and interactive imitation learning (e.g., DAgger)
- Build and maintain the data pipelines that support post-training, including teleoperation and demonstration collection, retargeting, dataset aggregation, and curation
- Drive the development cycle from simulation and offline evaluation to robustly transferring and fine-tuning policies on the robot
- Develop visuomotor policies for grasping, in-hand manipulation, and whole-arm manipulation, with reliable, low-latency execution on hardware
- Define evaluation metrics and regression tests for policy performance in simulation and on hardware, and use the results to guide technical decisions
- Collaborate closely with the simulation, data, hardware, and systems teams to diagnose system-level issues and co-develop solutions
- Maintain high standards for the codebase through testing, documentation, and peer code reviews
- Learning-Based Control: Hands-on experience training and debugging learned policies, including post-training methods such as RL fine-tuning, distillation, and imitation learning (e.g., DAgger)
- Data & Evaluation: Experience working with demonstration and teleoperation data, and building evaluation pipelines for learned policies
- Manipulation: Strong experience in robotic manipulation, including grasping, in-hand or whole-arm control, and high-DOF motion, with a track record of hardware deployment
- Software Engineering: Proficiency in Python, C++, and PyTorch or JAX, with experience building real-time robotic software stacks
- An MS with 3+ years of relevant experience, or a recent PhD, in Robotics, Computer Science, Electrical Engineering, or a related field
- Simulation Environments: Experience with physics engines such as IsaacSim, MuJoCo, or Drake for policy training and validation
- A proven track record of taking learned policies from a research/simulation environment and successfully deploying them on physical hardware
- Teleoperation: Experience with VR/haptic interfaces and retargeting algorithms for human-in-the-loop control
- Tactile Sensing: Experience integrating tactile/haptic feedback into manipulation pipelines
- Relevant experience specifically focused on learning-based manipulation or a closely related area
- Computer Vision: Familiarity with 6D pose estimation, point cloud processing, or visual-servoing
- Generalist Policies: Experience adapting large, general-purpose manipulation policies to specific tasks and embodiments
Skills
Sign in to read and join the discussion.