YongdongWang/dart_llm_tasks

DART-LLM Tasks Dataset

Description

DART-LLM Tasks is a dataset designed for evaluating language models in robotic task planning and coordination through few-shot learning. It contains 102 natural language instructions paired with their corresponding structured task decompositions and execution plans.

Dataset Structure

• Total examples: 102
• Complexity levels:
  • L1 (Basic): 47 examples
  • L2 (Medium): 33 examples
  • L3 (Complex): 22 examples

Features

• task_id: Unique identifier for each instruction
• text: Natural language instruction
• output: Structured task decomposition and execution plan
  • tasks: List of sub-tasks
    • task: Task name/identifier
    • instruction_function: Function specification
      • name: Function name
      • robot_ids: Involved robots
      • dependencies: Task dependencies
      • object_keywords: Relevant objects/locations

Task Categories

1. Movement Tasks: Navigation to specific areas
2. Excavation Tasks: Digging operations
3. Loading/Unloading: Material transfer
4. Avoidance Tasks: Obstacle and hazard avoidance
5. Coordination Tasks: Multi-robot operations

Usage

# Using the dataset for few-shot evaluation
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("path/to/dart_llm_tasks")

# Access examples
print(dataset[0])  # First example