Datasets:

PickNikRobotics
/

kinova_gen3_cube_stack_sim

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - vision-language-action
  - imitation-learning
  - manipulation
configs:
  - config_name: default
    data_files: data/*/*.parquet

Kinova Gen3 cube stacking (simulated)

360 demonstrations of color-conditioned cube stacking on a simulated Kinova Gen3, generated in MuJoCo by a scripted inverse-kinematics oracle and published in LeRobot v3.0 format. A red, a green, and a blue cube sit on a table. Each episode has a prompt such as stack the red cube on the green cube, and the arm picks the named cube and places it on the named target. All six color orderings are present, 60 episodes each.

The reference checkpoint trained on this dataset is pi05_kinova_gen3_cube_stack_sim.


Episodes	360
Frames	59,312
Tasks	6
Robot	Kinova Gen3 7-DoF with a Robotiq 2F-85 gripper
Control rate	10 Hz
Format	LeRobot v3.0

feature	dtype	shape	contents
`observation.state`	float32	[8]	measured position of `joint_1`...`joint_7` and `robotiq_85_left_knuckle_joint`
`action`	float32	[8]	commanded position, same channel order
`observation.images.scene`	video	[480, 480, 3]	fixed view of the workspace
`observation.images.wrist`	video	[480, 480, 3]	wrist-mounted camera
`observation.images.overview`	video	[480, 480, 3]	top-down view

Video is H.264, yuv420p. Cube positions are randomized per episode within the arm's top-down reach.

meta/train_layouts.jsonl gives the starting cube pose for each episode, keyed by episode_index. meta/eval_layouts.jsonl gives the 150 held-out layouts used to evaluate the reference checkpoint. Both record cube positions and yaws in the order red, green, blue, with held and target indexing into that order.

License

Apache-2.0. The data is PickNik-generated: simulated trajectories and renders from our own MuJoCo scene.

Kinova Gen3 cube stacking (simulated)

Contents

License