shivakanthsujit/d3il_avoiding_vision_224

D3IL Avoiding Vision 224

This is the original D3IL Avoiding demonstration dataset converted to the LeRobot v2.1 format and augmented with state-aligned 224 x 224 RGB observations.

It contains all 96 original Avoiding demonstrations (7,305 training frames), covering the benchmark's 24 avoidance modes with four demonstrations per mode. The numeric demonstrations are preserved from the original D3IL pickle logs.

Dataset summary

Property	Value
Episodes	96
Frames	7,305
FPS	29
Tasks	1
State dimension	4
Action dimension	2
Cameras	2
Image resolution	224 x 224

The camera observations are available under:

• observation.images.bp_cam
• observation.images.inhand_cam

Load the public dataset directly with LeRobot:

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("shivakanthsujit/d3il_avoiding_vision_224")

The images were produced by restoring every logged robot pose into the original D3IL MuJoCo environment and rendering the benchmark cameras. Actions were not replayed and physics was not advanced, so the images remain aligned with the original recorded states.

action stores the original next absolute desired controller target. To train with D3IL's native delta action, subtract the current desired target (the first two values of observation.state) at the corresponding horizon step from action.

The metadata also retains source_file, source_episode_index, and source_fps for provenance. state_is_pad, action_is_pad, and camera_is_pad are included for compatibility with the multi-task converter; all features in this single-task dataset are valid and unpadded.

Source

• D3IL repository: https://github.com/ALRhub/d3il
• Project page: https://alrhub.github.io/d3il-website/
• Paper: Towards Diverse Behaviors: A Benchmark for Imitation Learning with Human Demonstrations, ICLR 2024.

Please cite the original D3IL work when using this dataset.