HiFi-UMI trains robot policies without teleoperation
HiFi-UMI is a high-fidelity data-production system that uses a head-mounted camera rig to achieve 3 mm tracking accuracy for training robot manipulation policies without physical teleoperation. The authors released HiFi-UMI-2K, a 2,000-hour demonstration dataset showing that human-only data matches real-robot teleoperation performance.
Eliminating the need for physical robot hardware during data collection is a massive paradigm shift that dramatically lowers the cost and barrier to scaling generalist robotics models.
* Achieves 3 mm end-effector accuracy and sub-40 microsecond sensor synchronization via a portable head-mounted rig.
* Enables zero-robot post-training, allowing policies trained purely on human demonstrations to deploy directly on physical robot hardware.
* Matches the performance of traditional in-domain teleoperation data across multiple model families, including VLA and WAM architectures.
* Releases HiFi-UMI-2K, a 2,000-hour open dataset across 480+ scenes to help democratize large-scale robot dataset training.
DISCOVERED
1d ago
2026-07-29
PUBLISHED
1d ago
2026-07-29
RELEVANCE
AUTHOR
_akhaliq