About THOR
THOR is an open-source robotics lab building shared standards for robotics data, models and hardware. It runs on Bittensor Subnet 112, where customers post labelling work with an acceptance rule, miners produce the labels, and validators decide what passes. Only accepted data returns to the customer.
We start with data because it sits beneath simulation, hardware, research and deployment. THOR Data, which operates as Talo Robotics, THOR Hardware and THOR China are the first three businesses. Talo Robotics is a human demonstration and labelling firm for research teams.
The opportunity
As a Research Engineer on THOR Data you will own how labelling quality is measured. That covers the rubrics customers post with their footage, the hidden control samples validators use to check miners, and the scoring that decides which output is accepted and paid.
The work sits between robotics research and a live market. Your choices set what a verified action segment, object track or hand pose means, and how quickly a new customer specification can go from draft to accepted data.
What you'll do
- Write and test acceptance rubrics for teleoperation, egocentric and human labelling work
- Build the control samples and agreement measures validators use to score miners
- Run experiments on how labelling quality affects downstream policy training
- Turn customer specifications into tasks miners can complete and validators can check
- Publish the specifications and tools that can be open, so other teams can reuse them
What we're looking for
- Experience in robotics, computer vision or machine learning, in research or production
- Hands-on work with robot or human video data: action segmentation, object tracking, hand pose or similar
- A working grasp of how label quality affects model training
- Strong Python and comfort building data pipelines and evaluation tools
- Clear written English; Mandarin is useful, as part of the work runs through Shenzhen
- Comfort owning a problem end to end in a small team