Datasets:

webxos
/

nexus-fpv

license: mit
pretty_name: NEXUS FPV Physics Dataset (forest, dataset / auto)
tags:
  - robotics
  - reinforcement-learning
  - drone
  - fpv
  - physics-simulation
  - synthetic
  - trajectory
task_categories:
  - reinforcement-learning
  - robotics
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: telemetry.csv

NEXUS FPV Physics Dataset Sampler by webXOS

Auto-generated small FPV drone flight-telemetry and reinforcement-learning for experience sample dataset captured directly in the browser by NEXUS FPV. Each row is one physics tick recorded while a drone flew through a waypoint course, either under manual control or the built-in PID auto-pilot.

Play the game and make your own datasets: https://webxos.itch.io/nexus-fpv or download it from the /gym/ folder of this repo.

Generator: NEXUS FPV Physics Dataset Generator
Export timestamp: 2026-07-17T08:34:04.112Z
Scene: forest
Game mode: dataset
Control source: auto
Total telemetry records: 170
Total RL experiences: 179
Recording duration: 12.3s
Epochs / laps completed: 0
Collisions: 0
Path length: 39m

Files

File	Description
`telemetry.csv`	Flat, tabular telemetry — one row per physics tick (default HF viewer config).
`telemetry.jsonl`	Same telemetry data, newline-delimited JSON.
`experiences.jsonl`	`(state, action, reward, next_state, done, q_values)` tuples sampled during auto-pilot flight, ready for offline RL / imitation learning.
`metadata.json`	Full run metadata, including neural-network hyperparameters and waypoint pattern.

`telemetry.csv` / `telemetry.jsonl` columns

Column
`timestamp`
`mode`
`scene`
`epoch`
`iteration`
`pos_x`
`pos_y`
`pos_z`
`vel_x`
`vel_y`
`vel_z`
`rot_pitch`
`rot_yaw`
`rot_roll`
`target_x`
`target_y`
`target_z`
`current_waypoint`
`waypoint_name`
`distance_to_target`
`pattern_progress`
`q_value`
`reward`
`cumulative_reward`
`loss`
`epsilon`
`collisions`
`path_length`
`camera_mode`
`boost`
`auto_mode`

Neural network configuration

Architecture: [8,24,9]
Learning rate: 0.001
Gamma: 0.95
Epsilon (final): 0.9323284674748037 (min 0.05, decay 0.998)
Action space size: 9
State vector size: 8

Loading

from datasets import load_dataset

ds = load_dataset("path/to/this/dataset")
print(ds["train"][0])

Intended use

Trajectory sample data is for training/evaluating drone flight-control policies, imitation learning, offline RL, and physics-based motion prediction. Data is synthetic, generated entirely from an in-browser physics simulation — no real-world flight or personal data is included.

License

MIT