Datasetpublic

aboguszewski/robomimic

robomimic Dataset Summary This dataset contains demonstrations of robosuite tasks, separated into two directories: images and dino3-embeddings.The demonstrations in both of the directories are exactly the same. They differ by the format in which they store camera observations. images stores them as images, where dino3-embeddings stores the embeddings of those images, outputed by the small variant of the DINOv3 model. The dataset uses robotview and eye_in_hand… See the full description on the dataset page: https://huggingface.co/datasets/aboguszewski/robomimic.

sourceHugging Facemitupdated 3mo agoView on Hugging Face

0likes103downloads

Dataset Card

robomimic

Dataset Summary

This dataset contains demonstrations of robosuite tasks, separated into two directories: images and dino3-embeddings. The demonstrations in both of the directories are exactly the same. They differ by the format in which they store camera observations. images stores them as images, where dino3-embeddings stores the embeddings of those images, outputed by the small variant of the DINOv3 model.

The dataset uses robotview and eyeinhand camera positions. It also records rewards.

Block lifting and nut assembly demonstrations were sourced from the robomimic dataset. Block stacking demonstrations are original and were collected via robosuite's data collection script on a keyboard. All IIWA arm demonstrations are also original and collected in the same way. The machine generated dataset for block lifting is originally much larger than the samples available here. The ones here are the last 200 from the original robomimic dataset.

DINOv3 is an open-weights image embedding model.

For more information about the original dataset see the robomimic docs.

Details

File	Robot	Policy	Rewards	Image Resolution	# of Demonstrations	Mean Trajectory Length (approx)
block-lifting/panda/ph	Panda arm	proficient human	dense	512x512	200	48
block-lifting/panda/mh-better	Panda arm	"better" human operator	dense	512x512	100	72
block-lifting/panda/mh-okay	Panda arm	"okay" human operator	dense	512x512	100	93
block-lifting/panda/mh-worse	Panda arm	"worse" human operator	dense	512x512	100	145
block-lifting/panda/mg	Panda arm	machine generated	dense	512x512	200	150
block-lifting/iiwa/kh	IIWA arm	keyboard human	dense	512x512	11	54
nut-assembly/panda/ph	Panda arm	proficient human	dense	512x512	200	150
nut-assembly/panda/mh-better	Panda arm	"better" human operator	dense	512x512	50	216
nut-assembly/panda/mh-okay	Panda arm	"okay" human operator	dense	512x512	100	264
nut-assembly/panda/mh-worse	Panda arm	"worse" human operator	dense	512x512	100	357
block-stacking/panda/kh	Panda arm	keyboard human	dense	512x512	10	99
block-stacking/iiwa/kh	IIWA arm	keyboard human	dense	512x512	11	72

DINOv3 embeddings are 384-dimensional.

For more information about the tasks and environment see:

For more information about the robots see robosuite/manipulators.

Notes

—Mean trajectory length within a task is a decent metric of demonstration quality. As a rule of thumb ph ~ mh-better > mh-okay > mh-worse > mg.
—Machine generated trajectories are really weird. Not even remotely human-looking.
—Machine generated trajectories are not always succesful and when they are, they do not terminate after the success. They go on for 150 steps every time.
—Original robomimic datasets for nut assembly (square) had duplicated demonstrations. One of the operators' demonstrations from the mh-better dataset were also used for the ph dataset. They are deduplicated here, so that's the reason why nut assembly mh-better contains 50 demonstrations instead of a 100.

Sample Trajectories