HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone

Researchers introduced HiFi-UMI, a portable robot-free data-collection system designed to capture higher-fidelity manipulation demonstrations.

HiFi-UMI improves pose accuracy, synchronization, and camera coverage, yielding a dataset that is more suitable for training deployable policies.

Policies post-trained only on HiFi-UMI transferred directly to real robots and matched teleoperation performance across several model backbones.

Larger-scale pretraining on the same corpus further improved generalization and task success, including challenging precision manipulation tasks.

The results suggest real-world robot policies may be trainable with higher-quality robot-free demonstrations, reducing dependence on expensive robot data.