robo-semsim

A multi-source robotics image dataset for semantic safety evaluation and contextual scene annotation.

Dataset Structure

This dataset is published as source-specific subsets (configs):

• grand_tour
• ttpla
• bridge_data_v2_scripted

There is no combined subset in the current release workflow.

Schema

Each row contains:

• unique_id (string): deterministic ID built from dataset_source + relative_path
• image (image): image file
• mission (string): mission/context label
• camera (string): camera/view identifier
• original_filename (string): original image filename
• relative_path (string): relative path under source root
• dataset_source (string): source identifier (grand_tour, ttpla, bridge_data_v2_scripted)
• semantic_information (list[string]): contextual semantic annotations

Semantic Annotation Field

semantic_information is a flexible list of strings per image.

Current convention:

• semantic_information = [] means unannotated

Possible annotation prefixes:

• critical_object:
• unsafe_area:
• nominal_object:
• reasoning:

Example:

[
  "critical_object: moving forklift near pedestrian corridor",
  "unsafe_area: blind corner at corridor intersection",
  "nominal_object: marked walkway boundary",
  "reasoning: forklift path intersects pedestrian route within short stopping distance"
]

Mission Naming Conventions

• grand_tour: mission-specific label (human-readable + timestamp when available)
• ttpla: ttpla_drone_flying
• bridge_data_v2_scripted: bridge_manipulator_traj

Quick Start

  from datasets import load_dataset

  ds = load_dataset("aritrac-cub/robo-semsim", "ttpla", split="train")
  print(ds.features)
  print(ds[0]["unique_id"], ds[0]["semantic_information"])

Contribution Workflow (Annotations)

Contributions are welcome via Hugging Face Pull Requests.

Recommended process:

1. Load one subset.
2. Create/update annotations keyed by unique_id.
3. Keep image bytes unchanged; update metadata/annotation artifacts.
4. Open a PR to this dataset repo.

Example (upload annotation artifact as PR):

  from huggingface_hub import HfApi

  api = HfApi()
  api.upload_file(
      path_or_fileobj="annotations/ttpla_semantic_information.jsonl",
      path_in_repo="annotations/ttpla_semantic_information.jsonl",
      repo_id="aritrac-cub/robo-semsim",
      repo_type="dataset",
      create_pr=True,
      commit_message="Add semantic_information annotations for ttpla subset",
  )

Sources

• Grand Tour: https://grand-tour.leggedrobotics.com/
• TTPLA: https://github.com/R3ab/ttpla_dataset
• BridgeData V2: https://rail-berkeley.github.io/bridgedata/