Episodes Preview gr1 Visualizer
100 episodes · 50 fps

NVIDIA Arena GR1 Place Item and Close Door (TsFile)

Apache TsFile version of nvidia/Arena-GR1-Manipulation-PlaceItemCloseDoor-Task.

Overview

This dataset contains Isaac Lab trajectories for a GR1 humanoid manipulation task: place a ranch dressing bottle on the top shelf of a fridge, then close the fridge door. The converted data is the GR00T-LeRobot export of the 100 MimicGen-generated demonstrations described by the source dataset.

  • Episodes: 100
  • Frames: 32,735
  • Sampling rate: 50 fps
  • Split: train, episodes 0 through 99
  • Task: Place the ranch dressing bottle on the top shelf of the fridge, and close the fridge door.

The source repository also contains recorded, annotated, and generated HDF5 artifacts. Those HDF5 files and the RGB videos are not part of this numeric TsFile conversion. They remain available from the original dataset.

Schema (TsFile structure)

  • Time (INT64, milliseconds) is round(timestamp * 1000). The redundant source timestamp field is not stored separately.
  • episode_index and task_index are TAG columns. For example, select one trajectory with WHERE episode_index=0.
  • observation_state_0..25 (FLOAT) contains the 26 GR1 joint-state values.
  • action_0..25 (FLOAT) contains the 26 commanded joint positions.
  • annotation_human_action_task_description, frame_index, and sample_index (INT64) retain task/frame/sample identifiers; sample_index is the renamed source index field.
  • next_reward and observation_img_state_delta (DOUBLE), plus next_done (INT64), retain the remaining scalar trajectory fields.

Usage

Read data/arena_gr1_manipulation_place_item_close_door_task_train.tsfile with the Apache TsFile Java or Python SDK.

Camera videos are not uploaded here. See the original dataset's LeRobot videos.

Source & license

Downloads last month
30