Blame for README.md

---
license: cc-by-4.0
language:
- en
tags:
- neuroscience
- reinforcement-learning
- drosophila
- simulation
pretty_name: FlyControl paper data
---
# FlyControl data

Research records for **A Drosophila connectome prior enables adaptable embodied control**.

Code: [FlyControl](https://github.com/THU-NeuroML/FlyControl/tree/1a49c33bf1e6eab17cb2953123e02ecc2549a5f4).

| Figure | Experiment | Models |
|---|---|---|
| Figure 2 | Fly terrain locomotion | Adult, MLP and ER; 20 seeds each |
| Figure 3 | Fly turning control | Adult and MLP; 24 seeds each |
| Figure 4 | Fly speed control | Adult and MLP; 20 seeds each |
| Figure 5b–d, S7 | Ant flat ground and floor friction | Adult and MLP; 20 seeds each |
| Figure 5e | Ant U-Maze, Medium Maze and Large Maze | Seven controller families; 20 seeds per maze and family |
| Figure 5f–j | Ant Medium Maze trajectories and navigation | Adult and MLP; 20 seeds each |
| S3 | Developmental direction control | Adult and Larva; 6 seeds each |
| S4 | Developmental speed control | Adult and Larva; 6 seeds each |
| S5 | BANC descending and motor readouts | 6 seeds per readout; initial and final states |
| S6 | BANC regional recovery | 3 seeds per readout |
| S8 | Ant Large Maze trajectories, speed and route complexity | Seven controller families; 20 seeds each |
| Figure 6a–e | Cassie balance; Tripod, Go1 and Hexapod walking | Connectome and MLP; 20 seeds per body and family |
| Figure 6f | Cassie standing and lateral-motion stability | Connectome and MLP; 20 seeds each |
| Figure 6g | Go1 tilt, pace and heading stability | Connectome and MLP; 20 seeds each |
| S9 | Tripod direction and path stability | The Figure 6 Tripod policies |
| S10 | Hexapod path and lateral-corridor stability | The Figure 6 Hexapod policies |


The seven Ant maze controller families are Adult connectome, MLP, GRU, LSTM, Transformer, Reservoir and Reservoir+RO. The Ant flat-ground MLP uses two 1,378-unit ReLU hidden layers.

The archives contain episode records, trajectories, figure source tables, statistical summaries and experimental inputs. Corresponding policies are in [FlyControl-models](https://huggingface.co/UncleStrawberryBear/FlyControl-models).

## Figure 5, S7 and S8

`archives/figure5.tar.gz` contains:

- Ant flat ground: 4,000 evaluation episodes, motion endpoints and displacement trajectories from 40 policies.
- Floor friction: 4,000 episodes across four friction coefficients, with seed summaries for S7.
- Ant mazes: 42,000 evaluation episodes from 420 policies across U-Maze, Medium and Large mazes.
- Medium Maze: example trajectories, deadline curves, geodesic efficiency, route-turn groups and physical-failure records.
- Large Maze: seven example trajectories and speed records, route-stratified metrics and seed summaries for S8.
- Statistical source tables, including the 26- and 29-comparison families.
- Teacher-labelled training samples and an index mapping them to maze, controller, optimization seed and training stage.

Extracted files are under `data/figure5/`. Start with `experiment_map.csv`, `model_index.json`, `maze/episodes.csv`, `source_data/`, `S7/`, `S8/` and `training_labels_index.json`. Identical training-label files are stored once and referenced by checksum.

Maze training labels were generated using the GoalAnt teacher from [Farama-Antmaze-Dataset-Generation](https://github.com/LukaszSawala/Farama-Antmaze-Dataset-Generation/tree/fb253c9efbe2fa5ec256a6fc3f983ba33192d706), checkpoint SHA-256 `1b835cc5185756c5af8cd0720dae0e4895efcec48ee577c450d4a5a633fb953e`.

Optimization seed identifies the policy replicate; reset seed identifies an evaluation episode. Trajectory distances are in metres, time in seconds and velocity in metres per second. Each source table retains its original field names and units.

## Figure 6, S9 and S10

`archives/figure6.tar.gz` contains 22,400 evaluation episodes and their trajectories from 160 policies: four bodies, two controllers, 20 training seeds, seven conditions and 20 matched evaluation states per seed and condition.

- Figure 6a–e: Cassie balance and Tripod, Go1 and Hexapod walking; nominal trials, lateral pushes, reduced friction, 80-ms policy delay, 25% sensor loss and 60-s trials.
- Figure 6f: Cassie standing fraction, early lateral motion, reduced-friction standing duration and the matched snapshot trajectories.
- Figure 6g: Go1 tilt, pace, heading and the matched snapshot trajectories.
- S9: Tripod lateral displacement, heading error, late drift, outcome composition and success by condition.
- S10: Hexapod lateral displacement, late drift, corridor retention, first-exit times and success by condition.

Files are under `data/figure6/`. `experiment_map.csv` maps panels to source tables; `model_index.json` identifies the corresponding models. `twenty_seed/episodes.csv` contains the cohort, `details/` and `path_summary/` contain the main-figure source data, and `comprehensive/` contains the shared statistical and time-series data for Figure 6, S9 and S10, including the 10-endpoint and 3,150-comparison statistical families. These three analysis folders are under `twenty_seed/`.

Raw records use `raw/<body>/<method>_seed<training_seed>/<evaluation_seed>/<condition>.json` and `.npz`. The NPZ `trajectory` array contains time, qpos, qvel, executed Gaussian action, forward velocity, torso tilt and applied lateral force. Each episode JSON supplies `nq`, `nv` and `nu` for the variable-width blocks. Time is in seconds, distances in metres, velocity in metres per second, tilt in degrees and force in newtons. `experiment_protocol.json` lists the bodies, commands, training budgets, seeds and condition durations.

To download this group, use `manifests/figure6.jsonl` and `archives/figure6.tar.gz` in the example below, then extract with `--figure figure6`.

## Download and extract

```python
from huggingface_hub import snapshot_download
data = snapshot_download(
    'UncleStrawberryBear/FlyControl-data', repo_type='dataset',
    allow_patterns=['README.md', 'archive_index.json', 'extract_data.py',
                    'manifests/figure5.jsonl', 'archives/figure5.tar.gz'])
```

```bash
python /path/to/data-snapshot/extract_data.py --figure figure5 --output ./paper-data
```

Use `figure2`, `figure3`, `figure4`, `S3`, `S4`, `S5`, `S6` or `figure6` to extract the other archives. `archive_index.json` lists archive sizes and checksums; `manifests/` lists member paths, source paths and SHA-256 values. The extractor verifies these checksums. `figure_map.csv` maps paper figures to their archived records. Source and licensing information is in `RIGHTS.md`.

## License

The authors' original research data, trajectories, figure source tables and accompanying documentation are released under [Creative Commons Attribution 4.0 International (CC BY 4.0)](LICENSE). Third-party materials retain their upstream licenses and attribution; see [RIGHTS.md](RIGHTS.md).

Python utilities in this data repository use the [MIT License](LICENSE-CODE).