EvenFlow / README.md
yoichims's picture
Update README.md
307a024 verified
metadata
pretty_name: EvenFlow Benchmark
license: other
task_categories:
  - robotics
  - other
tags:
  - navigation
  - human-robot-interaction
  - trajectory-data
  - multi-agent
  - benchmarking
annotations_creators:
  - no-annotation
source_datasets:
  - original
language:
  - en
size_categories:
  - 1K<n<10K

EvenFlow Benchmark Dataset

EvenFlow is an evaluation suite for shared-space navigation, built from real-world human trajectory data.

Most benchmarks evaluate whether an agent can navigate around people.
EvenFlow evaluates whether an agent can navigate with them.

It converts real-world human trajectories into executable navigation tasks, enabling evaluation of coordination, timing, and interaction—not just collision avoidance.


Dataset Download.

git clone https://huggingface.co/datasets/standard-cognition/EvenFlow

What This Dataset Provides

The dataset consists of:

  • Layouts: static environment geometry
  • Scenes: human trajectory data over time
  • Tasks: executable navigation problems derived from real behavior
  • Tracks: time-indexed human trajectories

These components are structured to support trajectory-level evaluation of navigation planners.


Dataset Structure

benchmark/
  aligned_flow/
    tasks/
    scenes/
    layouts/
  cross_flow/
    tasks/
    scenes/
    layouts/
  interaction_constrained/
    tasks/
    scenes/
    layouts/

Each scenario family captures a different navigation regime:

  • Aligned Flow: motion aligned with surrounding traffic
  • Cross Flow: traversal across moving streams
  • Interaction-Constrained: navigation shaped by local human interactions

How to Use This Dataset

EvenFlow is designed for executable evaluation, not just analysis.

Typical workflow:

  1. Load a task (task.json)
  2. Resolve its scene and associated human trajectories
  3. Run a planner to generate a time-parameterized trajectory
  4. Evaluate the result using EvenFlow metrics

Code and evaluation tools:

👉 https://github.com/standard-ai/evenflow-benchmark


Responsible AI Considerations

Data Collection

Data was collected in real-world environments using overhead camera systems.
The dataset reflects naturally occurring human behavior in shared spaces.

Privacy

  • No raw video is released
  • No biometric identifiers are included
  • No personally identifiable information (PII) is present

All released data consists of anonymized trajectory representations.

Intended Use

This dataset is intended for research in:

  • robot navigation in human environments
  • multi-agent coordination
  • trajectory-based evaluation
  • human-aware motion planning

Limitations

  • Data is collected from a single physical environment (v1 release)
  • No demographic or identity-related attributes are included
  • Evaluation is performed offline (no closed-loop interaction)

We view this dataset as a foundation for broader multi-environment and interactive benchmarks.


License

This dataset is released under a custom research license.

  • Free for research and academic use
  • Commercial use requires a separate agreement

See the LICENSE file for full terms.