File size: 2,992 Bytes
47747ae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
---
license: mit
task_categories:
  - robotics
tags:
  - drone
  - ballistics
  - ball-drop
  - simulation
  - domain-randomization
pretty_name: HoverDrop Ballistic Drop Dataset
size_categories:
  - 1K<n<10K
---

# HoverDrop Ballistic Drop Dataset

Simulated point-mass ball drops for the **ballistics** half of LeDrone Track B
([hoverdrop](https://github.com/edgarmoreaualix/LeDrone)). Each record is one
drop: the conditions a release is solved from, the resulting landing, and a
**top-down camera frame** rendered at the release pose.

- **5,000 drops** total — Train: 4,000
  (16 shards) · Val: 500
  (2) · Test: 500
  (2)
- Frames: `128x128x3` uint8, one per drop.

## How it was generated

A point mass with quadratic air drag in wind-relative air is integrated
(semi-implicit Euler, 200 Hz) from a release state down to a flat ground plane.
Conditions are randomized per drop: altitude AGL `U[5,40] m`; 40% exact hover and
the rest moving (`U[0.5,12] m/s` horizontal at a random heading + `U[-2,2] m/s`
vertical); wind `U[0,8] m/s` horizontal at a random heading + `U[-0.5,0.5]`
vertical. The ball's mass (±10%) and drag coefficient (±20%) are randomized but
**not recorded in the model inputs** — this unobserved variation sets the
irreducible landing-error floor. The top-down frame is a nadir pinhole render
over a procedural `trajplanner` terrain map (seed varies per drop). Sim only; no
real flight data.

## Schema

Compressed `.npz` shards `{split}_{id}.npz`, one row per drop, deterministic
80/10/10 split by shard id:

| array         | shape          | dtype   | meaning |
|---------------|----------------|---------|---------|
| `inputs`      | `(n, 7)`       | float32 | `[altitude_agl, vx,vy,vz, wx,wy,wz]` (DropNet inputs) |
| `label`       | `(n, 3)`       | float32 | `[dx, dy, fall_time]` landing offset + fall time |
| `traj`        | `(n, L, 3)`    | float32 | 20 Hz ball trajectory (padded) |
| `traj_len`    | `(n,)`         | int32   | valid trajectory length |
| `ball`        | `(n, 3)`       | float32 | TRUE ball params `[mass, cd, radius]` (unobserved) |
| `drone_state` | `(n, 18)`      | float32 | release pose the frame is rendered from |
| `target_xyz`  | `(n, 3)`       | float32 | true landing point (drawn disc) |
| `map_seed`    | `(n,)`         | int32   | synthetic terrain seed |
| `frames`      | `(n, 128, 128, 3)` | uint8 | top-down render at release |

## Usage

```bash
pip install "hoverdrop[hf] @ git+https://github.com/edgarmoreaualix/LeDrone.git#subdirectory=hoverdrop"
python scripts/download_dataset_hf.py --repo-id Ethgar/hoverdrop-drops-5k --out-dir data/drops5k
```

```python
from hoverdrop.drops import DropDataset
ds = DropDataset("data/drops5k", split="train", load_frames=False)  # DropNet training
x, y = ds[0]        # inputs (7,), label (3,)
rec = DropDataset("data/drops5k", "test", load_frames=True).record(0)  # + frame
```

## License

MIT. Independent clean-room simulation; not affiliated with any third party.