ram-shreyas-naik-sabavat/SidewalkPilot_v1_and_v2

SidewalkPilot Series 1 and 2 Steering Dataset

SidewalkPilot Series 1 and 2 is the finalized camera-to-steering dataset for the baseline and failure/iteration model series. The dataset pairs real field images with steering servo labels in degrees, so a model can learn to map a camera frame to a steering command.

CARLA-assisted. The Series 1/2 models were trained on a blend of these real field images plus CARLA synthetic frames (down-weighted vs real). This repository holds the real labeled images; the CARLA synthetic set is published separately as SidewalkPilot_carla.

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_1_and_2
Hugging Face dataset	https://huggingface.co/datasets/ram-shreyas-naik-sabavat/SidewalkPilot_v1_and_v2
Hugging Face model namespace	https://huggingface.co/ram-shreyas-naik-sabavat

Dataset Contents

File or folder	What it contains
sidewalkpilot_v1_and_v2_dataset.tar	Clean archive containing the 2,224 JPG field images under sidewalkpilot_dataset/
steering_corrections.json	Steering labels, source names, and repeat weights for each labeled image

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 image archive before training or evaluation:

tar -xf sidewalkpilot_v1_and_v2_dataset.tar

This creates sidewalkpilot_dataset/ beside the metadata files. The archive excludes macOS metadata and upload-cache files.

Current Size

Item	Count
JPG images	2,224
Steering label entries	2,224
Label sources	13
Steering range	0 to 180 degrees

Label Format

steering_corrections.json is a JSON list. Each entry points to one image and stores the steering target used for training.

Field	Type	Meaning
image	string	Relative image path used by the training code
steering	number	Servo angle label in degrees
repeat	integer	Training repeat/weight value for that sample
source	string	Dataset source or field-test group for the image

Example entry:

{
  "image": "sidewalkpilot_dataset/photo_20260425_145756.jpg",
  "steering": 110.0,
  "repeat": 50,
  "source": "D0425_street_test"
}

Steering Label Meaning

The steering label is a servo angle in degrees.

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

Steering Distribution

Steering bucket	Count
0-45 hard left	69
45-75 left	128
75-85 soft left	281
85-95 straight	678
95-105 soft right	547
105-135 right	311
135-180 hard right	199

Source Breakdown

Source	Count	Purpose
D0328_first_dataset_relabel	315	First dataset relabel
D0329_first_dataset_relabel	413	First dataset relabel
D0425_street_test	65	Street test images
D0426_curves_shadows	53	Curves and shadow cases
D0427_curved_curb	72	Curved curb behavior
D0429_driveway_shadow_fix	53	Driveway and shadow cases
D0502_shadow_fix	154	Shadow robustness
D0502_19_hard_turn_curb_smoothness_fix	156	Hard turns, curb hugging, and smoothness
D0503_harsh_sidewalk	159	Harsh sidewalk surface cases
D0506_8pm_sidewalk	24	Evening / low-light sidewalk cases
D0510_v2_3_run_1	167	SidewalkPilot v2.3 field-run capture, run 1
D0510_v2_3_run_2	8	SidewalkPilot v2.3 field-run capture, run 2
D0510_v2_3_run_3	585	SidewalkPilot v2.3 field-run capture, run 3

Image Sizes

Resolution	Count
1280 x 720	1,496
1920 x 1080	413
320 x 240	315

The training pipeline resizes images before inference/training, so mixed capture resolutions are expected.

Basic Loading Example

from pathlib import Path
import json

dataset_root = Path("sidewalkpilot_dataset")
labels = json.loads(Path("steering_corrections.json").read_text())

first = labels[0]
image_name = Path(first["image"]).name
image_path = dataset_root / image_name
steering_degrees = float(first["steering"])

print(image_path, steering_degrees)

Training Use

The labels are intended for the SidewalkPilot steering trainer. Trainer code is maintained in GitHub, not duplicated in this dataset repository. The current training setup uses the image folder plus steering_corrections.json as the correction/label source.

Typical local training flow:

cd code/ai_models_datasets/series_1_and_2
python3 sidewalkpilot_trainer.py \
  --roots sidewalkpilot_dataset \
  --corrections steering_corrections.json \
  --model-version 2.4

Exact training commands may differ depending on whether CARLA data, source weighting, shadow augmentation, or other augmentation settings are being used.

Evaluation Use

The dataset is used to compare SidewalkPilot model checkpoints on the same labeled image set. Common metrics include:

Metric	Meaning
MAE	Mean absolute steering error in degrees
Median AE	Median absolute steering error in degrees
Max AE	Largest steering error in degrees
Signed Error	Directional bias of model predictions
Within 2 / 5 / 10 / 20 degrees	Count of predictions inside each tolerance band
Subset MAE	MAE grouped by field-test source

Intended Scope

This dataset supports the closed Series 1.x and 2.x research cycle. Series 1.x was the baseline working series, while Series 2.x pushed the same steering-only architecture to its limits and recorded the failure/iteration data used to design Series 3.x.

This dataset is finalized for the Series 1/2 Hugging Face release. New steering+throttle data should go into the separate Series 3 dataset instead.