README.md

metadata

task_categories:
  - robotics
dataset_info:
  features:
    - name: id
      dtype: string
    - name: subtask
      dtype: string
    - name: orientation
      dtype: string
    - name: target_orientation
      dtype: string
    - name: distance
      dtype: string
    - name: history
      dtype: string
    - name: current_view
      dtype: image
    - name: expected_view
      dtype: image
    - name: action
      dtype:
        class_label:
          names:
            '0': Subtask completed
            '1': Move forward
            '2': Turn left
            '3': Turn right
    - name: plan
      dtype: string
  splits:
    - name: train
      num_bytes: 3548909694.190918
      num_examples: 2980
    - name: validation
      num_bytes: 395613316.6690821
      num_examples: 332
  download_size: 3767539562
  dataset_size: 3944523010.86
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*

DeliveryBench

Project Page | Paper | Code

DeliveryBench is a city-scale embodied benchmark grounded in the real-world profession of food delivery. It is designed to evaluate the long-horizon planning and constraint-aware decision-making capabilities of LLM and VLM-based agents in realistic, procedurally generated 3D environments.

Dataset Summary

Agents in DeliveryBench must navigate procedurally generated 3D cities to maximize net profit while managing diverse constraints, such as delivery deadlines, transportation expenses, vehicle battery, and interactions with other couriers and customers. The environment includes:

• Diverse Road Networks and Buildings: Multiple cities with functional locations and various transportation modes.
• Realistic Resource Dynamics: Systematic evaluation of constraint-aware planning in a realistic, resource-dense environment.

Dataset Structure

The dataset contains trajectories of agents navigating these environments, including:

• current_view: Visual observation from the agent's perspective.
• expected_view: The target visual state for the subtask.
• plan: Natural language reasoning or planning steps associated with the trajectory.
• action: Discrete actions including moving forward, turning, or completing a subtask.
• metadata: Information regarding orientation, distance, and subtask history.

Citation

If you use this dataset in your research, please cite:

@article{deliverybench2025,
  title={DeliveryBench: Can Agents Earn Profit in Real World?},
  author={Authors list not provided},
  journal={arXiv preprint arXiv:2512.19234},
  year={2025}
}