Robot Learning Engineer
About the role
Robot learning is short of data with force in it. Most demonstrations are video and joint positions. The contact, force, hesitation and recovery that decide whether a task succeeds are missing. Our exoskeleton teleoperation captures all of it, because the operator feels what the robot touches.
You'll make our system the best data collection tool a robot learning team can buy. You'll adapt the hardware and software for collection, run collection campaigns and studies, and work directly with robot learning companies and labs as customers.
What you'll do
- Work with robot learning companies and research labs to understand their data needs. Adapt our system to fit their robot platforms, sensors, formats and throughput targets.
- Own the data collection pipeline end to end: logging, synchronization, calibration, metadata, quality checks, and export to the formats learning teams use (such as LeRobot and RLDS).
- Plan and run data collection campaigns, from protocol design and operator training to delivery.
- Show what our data is worth. Train baseline policies on it and run ablations, for example with and without force.
- Build tools that cut the cost of each useful episode: task labeling, episode segmentation, intervention and recovery tagging, and operator dashboards.
- Design and run studies on operator performance and data quality with our human factors team.
- Write up results for customers, papers and talks.
- Feed what you learn back to the controls, software and hardware teams.
What we're looking for
- MS or PhD in robotics, machine learning, computer science or a related field, or equivalent experience
- Hands-on experience with 6 or 7 DoF robot arms, ideally YAM arms
- Imitation learning experience: you've trained policies on real-robot demonstrations and gotten them working on hardware
- Experience collecting real-robot data through teleoperation and building the pipelines behind it
- Strong Python and PyTorch, and familiarity with ROS/ROS2 and robot data formats
- Experimental rigor: study design, ablations and statistics
- Comfort working directly with external customers and researchers, and explaining tradeoffs clearly
- A track record of debugging real hardware, not just simulation
Nice to have
We don't expect any one candidate to have all of these.
- Fine-tuning VLAs such as Pi0.7 or similar models on your own robot data
- Publications at CoRL, RSS, ICRA, NeurIPS or similar
- Experience using force/torque or tactile sensing in learning
- Experience on the data side at a robot learning company or lab
- Experience running data operations at scale, including operator teams and QA
- Experience with exoskeletons, haptics or teleoperation interfaces
About Mundane
Mundane builds force-feedback teleoperation for dangerous, hands-on work. A skilled tradesperson wears our exoskeleton and works through a robot on site, seeing in stereo and feeling the forces it meets, from outside the hazard. We're starting with energized electrical work and expanding across the trades. Every job also produces the force-rich data that robot learning is missing. We're a small, venture-backed team in Palo Alto.
Work authorization
Please tell us whether you're authorized to work in the US and whether you'll need visa sponsorship now or in the future.
Equal opportunity
Mundane is an equal opportunity employer. We welcome applicants from every background and don't discriminate on the basis of any legally protected characteristic.