bryandts/robotwin-action-prediction-dataset

Robotic Action Prediction Dataset

Dataset Description

This dataset contains triplets of (current observation, action instruction, future observation) for training models to predict future frames of robotic actions.

Dataset Structure

Data Fields

• current_frame: Input image (RGB) of the current observation
• instruction: Textual description of the action to perform
• future_frame: Target image (RGB) showing the expected outcome 50 frames later

Data Splits

The dataset contains:

• Total samples: ~91,303 (300 episodes × 304 frames average per episode)
• Tasks:
  • block_hammer_beat: "beat the block with the hammer"
  • block_handover: "handover the blocks"
  • blocks_stack_easy: "stack blocks"

Dataset Statistics

Task	Episodes	Frames per Episode
block_hammer_beat	100	200-300
block_handover	100	400-500
blocks_stack_easy	100	400-500

Dataset Creation

Source Data

• Simulation Environment: RoboTwin
• Image Resolution: At least 128×128 pixels
• Frame Offset: 50 frames between input and target