The ArmnetBench v0.1 core benchmark contains 50 human-teleoperated reference
trajectories per task plus 2,518 evaluation trajectories from 7 policies trained or
fine-tuned on those reference datasets, across 8 single-arm and 4 bimanual tasks performed
on the low-cost SO-101 robot arm. It was collected using the
Armnet arm farm.
The full release contains 3,718 labelled episodes: the 3,118 core benchmark
episodes plus 600 episodes from extra runs. The extra runs… See the full description on the dataset page:
https://huggingface.co/datasets/armnet/armnetbench_v01_robometer.
