Founding MTS - Dexterous Manipulation (Robot Learning)

Stealth AI Startup; Redwood City (CA); On-site; USD 220,000 - 300,000; Full time

Job summary

Stealth AI Startup in Redwood City is seeking a researcher to own how robots learn dexterous manipulation, from data collection to policy software running on hardware.

You will build teleoperation pipelines, retarget human motion to robot hands, train policies, and deploy them on real arms, with O(1) sponsorship and equity options. This role focuses on hands-on hardware and real-world evaluation.

Qualifications

• Experience training manipulation policies on real robots with measurable success.
• Hands-on with high-DoF manipulation, not only parallel grippers.
• Built or deeply used a teleoperation or data collection system.

Responsibilities

• Build teleoperation and data-collection pipelines with dexterous hands.
• Retarget human motion to robot hands and select demonstrations.
• Train and deploy dexterous policies on real hardware.
• Diagnose policy failures on hardware and fix data/models.
• Collaborate with world/model teams to turn video into skills.

Skills

Manipulation policy

Teleoperation systems

Data collection

Imitation learning

RL fine-tuning

Job description

Role Description

Dexterous Manipulation. You own how our robots get good with their hands, from the data we collect to the policy that runs on the robot. You decide what to capture, how to teleoperate, and how to learn from it, and you measure success on real hardware.

You will:

• Build and run our teleoperation and data collection: gloves, hand and wrist tracking, low-latency control of dexterous hands on real arms.
• Retarget human hand motion to robot hands, and decide which demonstrations are worth collecting.
• Train and post-train dexterous policies (imitation learning, diffusion or flow-matching policies, VLAs, RL fine-tuning) and deploy them on the robot.
• Close the loop: find where the policy fails on hardware, trace it back to data or model, and fix it.
• Work with our human modeling and world model teams to turn human video into robot skills.

Your first project: take one contact-rich bimanual task from first teleop demo to a policy that succeeds reliably on our real dexterous hands, with the numbers to show it.

You work with real bimanual arms, dexterous hands and teleoperation rigs. Compute is a cluster of B300s. Cash is $220K to $300K depending on experience, plus founding-tier equity. O-1 and H-1B sponsorship.

Qualifications

• You have trained a manipulation policy that works on a real robot, and can show the success rate.
• Hands-on with dexterous hands or high-DoF manipulation, not only parallel grippers.
• You have built or deeply used a teleoperation or data collection system, and know what makes demonstrations good or bad.
• You have carried a full loop alone: data, training, deployment, evaluation.
• You want the robot to do something hard with its hands.

Nice to have

• Gloves, motion capture, or VR teleop (MANUS, Rokoko, Vive, Quest, Apple Vision Pro).
• Tactile sensing, force control, or contact-rich in-hand manipulation.
• Human-to-robot retargeting, or learning from human video.
• Using LLMs or agents to scale data, evaluation, or research loops. (Pure RAG or chatbot work is less relevant here.)

A PhD helps but is not required. We care most about what you have built.