ram-shreyas-naik-sabavat/SidewalkPilot_v3_and_v4

SidewalkPilot v3 and v4 Steering Dataset

This is the shared real-world dataset for SidewalkPilot Series 3 and the experimental Series 4 temporal models. Every image is paired with a logical steering-servo target in degrees and the physical throttle command recorded at capture time.

Series 3 learns the current target from one image. Series 4 keeps the same images, labels, augmentation policy, and frozen time-window split while testing whether previous steering targets and future-target supervision improve steering. Series 4 removes throttle from the model output because about 95% of this dataset is full throttle; the throttle label remains published as raw evidence and for future dataset work.

Aggregate (2026-07-02 through 2026-07-12): 81,237 manually-driven sidewalk frames captured across five runs, decode-verified and curated. All labels are raw human stick commands (imitation learning), never autonomous/model-predicted — Series 3 must not be seeded with old Series 2.x model-predicted labels.

Not CARLA-assisted. Series 3 and 4 models use real RC-car photos only. The separate SidewalkPilot_carla repository is not mixed into this dataset.

Project code and documentation are maintained in the GitHub repo:

Resource	Link
GitHub repository	https://github.com/RamCodesBetter/SidewalkPilot
Training code	https://github.com/RamCodesBetter/SidewalkPilot/tree/lidar-aeb-v2/code/ai_models_datasets/series_3_and_4
Hugging Face dataset	https://huggingface.co/datasets/ram-shreyas-naik-sabavat/SidewalkPilot_v3_and_v4
Hugging Face model namespace	https://huggingface.co/ram-shreyas-naik-sabavat

Dataset Contents

File or folder	What it contains
sidewalkpilot_v3_and_v4_dataset.tar	Archive containing the shared Series 3/4 JPG field dataset
labels.json	Dict-style image -> {steering, throttle} for every frame (the sole label file)

Trainer source and tests are versioned only in GitHub so there is one canonical code history. The dataset repository contains data and its card, not duplicate executable code.

Extract the archive and place the root label manifest beside the images before training:

tar -xf sidewalkpilot_v3_and_v4_dataset.tar
cp labels.json sidewalkpilot_dataset/labels.json

Current Size

Item	Count
JPG images	81,237
Steering/throttle label entries	81,237
Label sources	5 human-driving runs (2026-07-02 through 2026-07-12)
Steering range	0 to 180 degrees (logical; 90 = straight)
Throttle range	0.00 to 1.00

Label Format

The repository publishes labels.json at its root. After extraction, copy it to sidewalkpilot_dataset/labels.json; that is the file loaded by --roots sidewalkpilot_dataset.

Field	Type	Meaning
image filename key	string	Captured image filename
steering	number	Logical steering servo angle in degrees (0-180)
throttle	number	Forward motor command (0.00-1.00)

Example entry:

{
  "2026_07_02_run_2__photo_20260702_170623_008443.jpg": {
    "steering": 90,
    "throttle": 1.0
  }
}

labels.json is the single source of labels for this dataset; there is no separate corrections/override file.

Steering Label Meaning

The steering label is a logical servo angle in degrees.

Steering value	Meaning
0	Hard left
90	Straight / center
180	Hard right

Throttle Label Meaning

The throttle label is the forward motor command used by the car at the frame.

Throttle value	Meaning
0.00	Stop
1.00	Full forward

Reverse is not a Series 3 model output. Braking, stopping, and reverse behavior remain runtime/safety responsibilities.

Steering Distribution

The nine buckets below are the model's steering classes (STEER_CLASS_BINS).

Steering bucket	Count	Share
HL 0-45 hard left	691	0.9%
L 45-60 left	1,027	1.3%
L+ 60-75 left	2,683	3.3%
SL 75-85 soft left	4,199	5.2%
ST 85-95 straight	53,910	66.4%
SR 95-105 soft right	2,795	3.4%
R 105-120 right	3,899	4.8%
R+ 120-135 right	4,061	5.0%
HR 135-180 hard right	7,972	9.8%

Grouped: left 10.6%, straight 66.4%, right 23.1%. The set is center-heavy and skews right of center: the car has a mechanical left-pull, so straight driving needed a slight right hold. Horizontal-flip augmentation (mirror the image and set the new label to 180 - steering) symmetrizes left/right, and class-balanced sampling counters center dominance.

Throttle Distribution

Throttle bucket	Count	Share
1.00 full forward	77,082	94.9%
0.95-0.999	2,914	3.6%
0.50-0.95	494	0.6%
0.01-0.50	231	0.3%
0.00 stop	516	0.6%

Throttle is effectively constant (~95% at full) because these batches were driven flat-out. There is almost no throttle variance, so the throttle head cannot learn meaningful throttle control from this data alone — treat it as a steering dataset until varied-throttle runs are added.

Source Breakdown

Source	Count	Purpose
2026_07_02_run_1	6,685	Manual human driving, 2026-07-02 run 1 (crash-truncated tail removed)
2026_07_02_run_2	43,999	Manual human driving, 2026-07-02 run 2 (main capture)
2026_07_07_run_1	6,524	Manual human driving, 2026-07-07
2026_07_12_run_1	2,102	Manual human driving, 2026-07-12 run 1
2026_07_12_run_2	21,927	Manual human driving, 2026-07-12 run 2 (103-frame segment culled)

Image Sizes

Resolution	Count
1280 x 720	81,237

All Series 3 and 4 trainers resize images to 320x180 by default.

Data Quality and Known Limitations

• Full Pillow decode-verify pass; 154 empty crash-tail frames removed, 3,143 frames culled from selected time ranges. No corrupt/truncated frames remain.
• Throttle is near-constant (see Throttle Distribution) — not learnable from this batch; needs varied-throttle runs.
• Steering is center-heavy and right-skewed (see Steering Distribution) — use horizontal-flip augmentation + class-balanced sampling.
• Consecutive frames are nominally 10 Hz and near-duplicates; split train/val by time segment, not randomly, to avoid leakage inflating validation scores.
• Series 4 rejects temporal windows that cross photo runs, capture gaps, or train/validation boundaries.

Basic Loading Example

from pathlib import Path
import json

dataset_root = Path("sidewalkpilot_dataset")
labels = json.loads((dataset_root / "labels.json").read_text())  # dict: image -> {steering, throttle}

first_image = next(iter(labels))
image_path = dataset_root / first_image
steering_degrees = float(labels[first_image]["steering"])
throttle = float(labels[first_image]["throttle"])

print(image_path, steering_degrees, throttle)

Training Use

The commands below run from the linked GitHub checkout; trainer code is not stored in this Hugging Face dataset repository.

Series 3 uses one image to predict the current nine-class hybrid steering target and an optional throttle output. Its current head has 19 raw values: nine class logits, nine within-class offsets, and one throttle value.

cd code/ai_models_datasets/series_3_and_4
python3 series_3_sidewalkpilot_trainer.py \
  --roots sidewalkpilot_dataset \
  --model-version 3.0 \
  --epochs 25

Series 4 is experimental and steering-only. Each predicted horizon has 18 raw values: nine class logits and nine within-class offsets. The defaults use three previous or future targets, approximately 0.3 seconds at the labeled capture rate.

# PC: previous targets -> current target
python3 series_4_0pr_sidewalkpilot_trainer.py --epochs 25

# CF: image -> current target plus future-target supervision
python3 series_4_0fg_sidewalkpilot_trainer.py --epochs 25

# PCF: previous targets -> current target plus future-target supervision
python3 series_4_0ac_sidewalkpilot_trainer.py --epochs 25

Each run writes one regular/final checkpoint and one best-current-steering-MAE checkpoint:

Trainer	Regular/final	Best
4.0pr	4.0p	4.0r
4.0fg	4.0f	4.0g
4.0ac	4.0a	4.0c

Series 4 checkpoints are experimental. They are not added to the live controller until ONNX validation, Jetson latency testing, offline comparison against v3.4, and field testing pass.

Evaluation Use

Series 3 evaluation retains steering and throttle metrics. Series 4 ranks the current horizon using steering metrics and reports each future horizon separately. Common metrics include:

Metric	Meaning
Steering MAE	Mean absolute steering error in degrees
Throttle MAE	Series 3 only: mean absolute throttle-command error
Balanced-9 recall	Mean recall across the nine steering classes
Turn exact / +/-1	Exact and adjacent-class turn accuracy
Signed steering error	Directional steering bias
Signed throttle error	Over-driving or under-driving bias
Field subset metrics	Metrics grouped by route, lighting, source, or capture mode

Augmentation Preview

Use the test helper to preview the shared Series 3/4 augmentation variations before training:

python3 ../../test_files/preview_series3_augmentations.py \
  sidewalkpilot_dataset/<example>.jpg \
  --output /tmp/series3_augmentations.jpg

Intended Scope

This dataset supports the Jetson-only Series 3 production line and the parallel Series 4 temporal experiments. Series 4 changes the architecture and temporal supervision, not the underlying evidence, so comparisons remain controlled.

The dataset should be updated only with synchronized image, steering, and throttle labels from deliberate human field driving. CARLA data remains separate and must not be silently mixed into this real-only dataset.