Robotics Movement Object-Affordance Dataset
This repository contains the Hugging Face dataset named robotics-movement-object-affordance-v1 for robotics movement, object recognition, and object manipulation planning. The dataset is designed to help a robot or robotics model understand common objects, what they are, which parts can be used for grasping, and what kinds of movements are appropriate for moving them safely.
The dataset currently contains 20 samples. Each sample describes one real-world object category, such as a bottle, mug, spoon, book, ball, tool, food object, textile, furniture item, or container. The records include object names, categories, short descriptions, typical object parts, affordances, grasp candidates, placement candidates, safety notes, robotic action steps, and source attribution.
This dataset is intentionally not a collection of fake robot trajectories. It does not contain synthetic joint angles, fake camera images, fake point clouds, or copied demonstrations from another robotics dataset. Instead, it is a source-attributed object-affordance dataset. It combines public Wikipedia summaries, open-source-style robotics knowledge, and human-written manipulation templates. The goal is to provide grounded semantic information that can support object understanding and movement planning before real sensor or robot demonstration data is added.
The dataset is useful for models that need to reason about object manipulation. For example, a model can learn that a mug often has a handle, that a handle is a good grasp region, that the mug should be lifted upright, and that the base should be placed flat on a table. A bottle may need slower motion if it contains liquid. A tennis ball may roll or deform. A towel may need a different grasp strategy from a wooden block. These distinctions are represented as text labels and action templates rather than as fabricated trajectories.
The dataset was created from the ground up using public sources and open-source robotics concepts. It does not wholesale copy another dataset. Wikipedia is used for public object descriptions and source attribution. The manipulation fields are written as practical robotics templates based on general object affordance reasoning. If you later add real robot demonstrations, the recommended approach is to store them in a separate data folder with their own consent, license, citation, sensor description, robot platform, environment description, and task metadata.
The main dataset files are data/object_affordance.jsonl and data/object_affordance.parquet. The JSONL file is compatible with Hugging Face Datasets, and the Parquet file is the faster tabular option for analytics and downstream pipelines. Each JSONL line is one JSON object. Every record has an id, an object_name, a category, a description, a list of typical_parts, a list of affordances, a list of robotic_actions, a list of grasp_candidates, a list of placement_candidates, a list of safety_notes, and source fields including source_titles, source_urls, and source_license_notes.
The source object records are stored in data/source_objects.json. These are the base records used by the build script. The build script reads those records, fetches public Wikipedia summaries when available, and writes the final Hugging Face-compatible JSONL dataset. The script also writes cached summaries to data/wikipedia_summaries.json so repeated builds do not need to fetch the same summaries again.
To build the dataset, run the build script from the repository root. On Windows PowerShell, use this command:
cd "c:\Users\opent\OneDrive\Nete\robotics-movement-dataset"
python scripts/build_dataset.py
To validate the dataset, run the validation script:
python scripts/validate_dataset.py
The validator checks the schema, required fields, duplicate IDs, duplicate object names, duplicate full records, obvious fake or placeholder markers, missing source attribution, and basic usefulness requirements. If repeated action or affordance patterns are found, the validator prints a warning. Repeated patterns are allowed because different real-world objects can share valid manipulation patterns. For example, many containers can be lifted and placed, and many handled tools are grasped by the handle.
To install the Python dependencies, run:
pip install -r requirements.txt
To load the dataset locally with Hugging Face Datasets, use:
from datasets import load_dataset
dataset = load_dataset("c:\\Users\\opent\\OneDrive\\Nete\\robotics-movement-dataset")
print(dataset)
print(dataset["object_affordance"][0])
The dataset is licensed under CC-BY-4.0. Source materials remain under their original licenses. Wikipedia content is licensed under CC BY-SA, and the dataset records summarize public information rather than copying long passages. When publishing this dataset publicly, keep the source attribution fields and license notes so users can understand where the information came from.
This dataset is best understood as a semantic robotics dataset rather than a motion-capture dataset. It is suitable for object affordance classification, natural-language manipulation planning, grasp-region reasoning, placement reasoning, safety-rule generation, and dataset construction pipelines. It is not sufficient by itself for training low-level robot control policies that require real trajectories, force readings, tactile data, depth images, RGB images, or joint commands.
A recommended future extension is to add a real demonstration split only if you have lawful and ethical data. Real demonstration records should include the robot platform, end effector, environment, task description, object identity, sensor streams, timestamps, trajectory format, coordinate frames, action space, success label, failure modes, human consent status if people appear, and license or usage restrictions. Until real data is available, this repository should remain a non-synthetic semantic object-affordance dataset, not a fake trajectory dataset.