mihyun1115/piper_data

Piper Sim Pick-and-Place Multitask Dataset

Scripted-expert (inverse-kinematics) demonstrations for a 6-DOF Agilex Piper arm + gripper performing 10 distinct language-conditioned pick-and-place tasks in a MuJoCo simulation.

Tasks

Each episode is labeled with one of 10 fixed natural-language instructions:

task_id	instruction
0	put the red block in the basket
1	put the yellow block on the plate
2	put the red block on the plate next to the basket
3	put the yellow block in the basket next to the plate
4	put the red block in the basket and the yellow block on the plate
5	pick up the block on the plate and put it in the basket
6	pick up the block that is not on the plate and put it in the basket
7	pick up the block in the basket and put it on the plate
8	pick up the block that is not in the basket and put it on the plate
9	stack the red block on the yellow block

Scene objects: a red 5cm cube, a yellow 5cm cube, a basket (26x17x8cm), and a 15cm green plate. Object positions/rotations are randomized every episode within a validated-reachable workspace.

This release contains the first 50 episodes of each task (500 total, episode_0000-episode_0049).

Directory layout

task00/episode_0000.npz ... episode_0049.npz
task01/episode_0000.npz ... episode_0049.npz
...
task09/episode_0000.npz ... episode_0049.npz

Episode format

Each episode_NNNN.npz contains (T = number of control steps in that episode, varies per episode; control runs at 20Hz):

key	shape	dtype	description
front	(T, 240, 320, 3)	uint8	front camera RGB
wrist	(T, 240, 320, 3)	uint8	wrist camera RGB (mounted on link6)
state	(T, 7)	float32	joint1-6 angle (rad) + gripper (0=closed, 1=open)
action	(T, 7)	float32	target joint1-6 angle (rad) + gripper command (0-1)
instruction	scalar	str	one of the 10 instructions above
task_id	scalar	int	0-9, index into the instruction table above

Note for model training: task_id is a generation-time label, not meant to be fed to a model as input — doing so would let a model "cheat" by looking up a task number instead of actually reading instruction. For training/evaluating a language-conditioned policy, only instruction should be used as the language-conditioning signal.

Source

Collected with a MuJoCo IK scripted expert from the piper simulation project.