penpenzh/pocketvla-dataset-2D-grab

PocketVLA Dataset 🦾

Fully simulated dataset for PocketVLA — a 3M-parameter educational VLA. Generated by a built-in 2D simulator (seed=42, fully reproducible), no external data required.

Task: a 2-link planar arm performs language-guided pick-and-place in a top-down scene — 3 balls (red/green/blue) + 1 bin, randomized per episode. Given an instruction like "put the green ball into the bin", the model must grasp the matching ball and place it in the bin.

Files

File	Purpose	Size
train.npz	SFT training (expert demos)	7,000 episodes / 184,364 samples
val.npz	Validation (held-out episodes)	3,000 episodes / 79,150 samples
test.npz	Fixed closed-loop test set	500 scenarios
grpo.npz	GRPO RL training scenarios	3,000 scenarios
ood_test.npz	OOD generalization test (unseen distractor per scene)	300 scenarios

Fields

train / val (one sample per step):

Field	Shape	Description
images	(N, 64, 64, 3) uint8	Top-down observation (bin visible, target not marked)
tokens	(N, 10) int64	Instruction tokens (vocab 22, pad=0)
proprio	(N, 6) float32	(q1/π, q2/π, ee_x/1.1, ee_y/1.1, gripper, holding)
actions	(N, 3) float32	(Δq1, Δq2, grip), joints normalized by 0.15 rad
target_xy / bin_xy	(N, 2) float32	Ball/bin offset from EE (aux labels)
color_id / instruction / ep_id	—	Commanded color / instruction text / episode id

test / grpo / ood_test (one stored initial scenario per row): balls (n,3,2), bin_pos (n,2), q0 (n,2), color_id, tokens, instruction; ood_test adds distractor_colors / distractor_pos / distractor_shapes.

Usage

import numpy as np

data = dict(np.load("train.npz"))
print(data["images"].shape, data["actions"].shape)
# (184364, 64, 64, 3) (184364, 3)

Full training/evaluation pipeline (SFT → GRPO → closed-loop & OOD eval): PocketVLA.

Citation

@misc{pocketvla2026,
  title  = {PocketVLA: Building a 3M VLA from Absolute Zero},
  author = {Enming Zhang},
  year   = {2026},
  url    = {https://github.com/penpenzh/pocketvla}
}