This dataset was derived from poolvarine/robomimic-mh-lift-image-dense.
Applied preprocessing:
Trimmed each episode after the first next.done == true row (inclusive).
Shifted every next.reward by subtracting MAX_DENSE_REWARD = 0.5555555555555556.
Snapped numerically tiny values to exactly 0.0 using tolerance 1e-12.
Set terminal success rewards to exactly 5.
Updated metadata to reflect the new total frame count: 29915.
Removed trailing post-success rows: 1212.… See the full description on the dataset page:
https://huggingface.co/datasets/poolvarine/robomimic-mh-lift-image-dense-negshift-success5.
