Pantheon built an in-house operation to collect the most diverse UMI dataset in robotics. It scaled to over 1 million tasks in eight weeks. The average cost dropped to $10 per hour of data. Operators were paid three times the local living wage. A 100-hour annotated sample of the dataset is now public. The team used freeform and scripted collection methods. They also captured failure and recovery data. Hardware was iterated three times to cut production costs to $11 per gripper pair. Custom firmware and infrastructure supported the operation. The dataset aims to enable zero-shot task generalization in robotics.
What changed
Existing UMI datasets cost $60–150 per hour and lacked task diversity.
What it unlocks
Collecting task-diverse, environment-diverse UMI data at scale for robotics training.
- 1 million tasks
- $10 per hour of data
- 90 operators in 8 weeks
- 100-hour dataset sample
What you need to act on it
- custom UMI hardware
- in-house data collection operation
- Argus annotation pipeline
Sources