Episodes Preview LeRobot Visualizer
240 episodes · 10 fps · 1 camera · 96×96 av1

AR-VLA: True Autoregressive Action Expert for Vision-Language-Action Models

This dataset is used in the paper AR-VLA: True Autoregressive Action Expert for Vision-Language-Action Models.

Dataset Description

This dataset contains teleoperation trajectories for the PushT2 environment, a variant of the PushT task. It was used to evaluate AR-VLA, an autoregressive Action Expert that generates actions as a continuous causal sequence while maintaining history through a long-lived memory.

Usage

The dataset is compatible with the LeRobot library. To train an Auto Regressive Transformer (ART) policy on this dataset, you can use the following command (ensuring the external policy package lerobot_policy_art from the official repository is installed):

lerobot-train \
  --policy.discover_packages_path=lerobot_policy_art \
  --policy.type=art \
  --dataset.repo_id=you2who/pusht2-teleop-v30 \
  --env.type=pusht \
  --env.task=PushT2-v0 \
  --steps=200000 \
  --batch_size=64

Note: If you override action chunking or history settings, ensure that history_length + chunk_size is aligned with the dataset window.

Downloads last month
57

Paper for you2who/pusht2-teleop-v30