keivalya/nemotron-vla-metaworld

Nemotron-VLA MetaWorld Expert Demonstrations

Expert demonstration dataset for training Vision-Language-Action (VLA) models on MetaWorld robot manipulation tasks.

Built for the Nemotron-VLA project.

Dataset Summary

Stat	Value
Tasks	50 MetaWorld environments
Total transitions	187,252
Camera	corner2 (third-person view)
Expert policy	MetaWorld built-in scripted policies
Image format	PNG (compressed)

Data Fields

Field	Type	Description
image	bytes (PNG)	RGB camera observation
state	float32[39]	Robot proprioceptive state
action	float32[4]	Expert action (xyz + gripper)
instruction	string	Natural language task instruction
env_name	string	MetaWorld environment name

Tasks

Task	Instruction	Transitions	Episodes	Success
assembly-v3	insert the peg into the hole	4,372	50	100.0%
basketball-v3	place the ball into the basket	4,560	50	100.0%
bin-picking-v3	pick up the object from the bin	5,647	50	100.0%
box-close-v3	close the box by placing the lid on top	5,409	50	90.0%
button-press-topdown-v3	press the button from above	3,311	50	100.0%
button-press-topdown-wall-v3	press the button from above near the wall	2,880	50	100.0%
button-press-v3	press the button from the front	2,974	50	100.0%
button-press-wall-v3	press the button near the wall	3,295	50	98.0%
coffee-button-v3	press the coffee machine button	1,776	50	100.0%
coffee-pull-v3	pull the coffee mug toward you	3,962	50	100.0%
coffee-push-v3	push the coffee mug away	2,746	50	100.0%
dial-turn-v3	turn the dial to the target angle	3,609	50	100.0%
disassemble-v3	pull the peg out of the hole	5,516	50	64.0%
door-close-v3	close the door	3,353	50	100.0%
door-lock-v3	lock the door by rotating the lock	5,877	50	100.0%
door-open-v3	open the door by pushing the handle	4,185	50	96.0%
door-unlock-v3	unlock the door by rotating the lock	2,206	50	100.0%
drawer-close-v3	push the drawer closed	3,921	50	100.0%
drawer-open-v3	pull the drawer open	4,406	50	100.0%
faucet-close-v3	turn the faucet to close it	3,249	50	100.0%
faucet-open-v3	turn the faucet to open it	2,951	50	100.0%
hammer-v3	use the hammer to drive the nail in	3,292	50	100.0%
hand-insert-v3	insert the object into the target slot	2,939	50	100.0%
handle-press-side-v3	press the handle down from the side	2,142	50	100.0%
handle-press-v3	press the handle down	1,421	50	100.0%
handle-pull-side-v3	pull the handle from the side	3,969	50	100.0%
handle-pull-v3	pull the handle toward you	5,088	50	100.0%
lever-pull-v3	pull the lever down	3,932	50	100.0%
peg-insert-side-v3	insert the peg from the side	5,225	50	84.0%
peg-unplug-side-v3	unplug the peg from the side	5,046	50	96.0%
pick-out-of-hole-v3	pick the object out of the hole	5,839	50	100.0%
pick-place-v3	pick up the object and place it at the goal	2,635	50	100.0%
pick-place-wall-v3	pick up the object and place it over the wall	4,236	50	100.0%
plate-slide-back-side-v3	slide the plate back from the side	2,302	50	100.0%
plate-slide-back-v3	slide the plate back to the starting position	3,148	50	100.0%
plate-slide-side-v3	slide the plate sideways to the target	2,408	50	100.0%
plate-slide-v3	slide the plate to the target	2,497	50	100.0%
push-back-v3	push the object back to the starting position	5,346	50	70.0%
push-v3	push the object to the goal	3,042	50	100.0%
push-wall-v3	push the object to the goal near the wall	4,002	50	100.0%
reach-v3	reach to the target position	2,398	50	100.0%
reach-wall-v3	reach to the target position near the wall	2,298	50	100.0%
shelf-place-v3	place the object on the shelf	4,128	50	100.0%
soccer-v3	kick the soccer ball to the goal	3,782	50	98.0%
stick-pull-v3	use the stick to pull the object closer	6,375	50	58.0%
stick-push-v3	use the stick to push the object away	3,665	50	100.0%
sweep-into-v3	sweep the object into the hole	3,052	50	92.0%
sweep-v3	sweep the object to the goal	4,412	50	100.0%
window-close-v3	slide the window closed	4,034	50	100.0%
window-open-v3	slide the window open	4,394	50	100.0%

Quick Start

from datasets import load_dataset
from PIL import Image
import io

# Load full dataset
ds = load_dataset("keivalya/nemotron-vla-metaworld")

# Filter by task
push = ds["train"].filter(lambda x: x["env_name"] == "push-v3")

# View an image
img = Image.open(io.BytesIO(push[0]["image"]))
img.show()

# Get state and action
state = push[0]["state"]   # list of 39 floats
action = push[0]["action"] # list of 4 floats
instruction = push[0]["instruction"]  # "push the object to the goal"

Used By

• Nemotron-VLA — Vision-Language-Action model using NVIDIA RADIO + Nemotron Nano 9B

License

MIT. MetaWorld is MIT licensed.