ram-shreyas-naik-sabavat/SidewalkPilot_v3

SidewalkPilot Series 3 Steering and Throttle Dataset

SidewalkPilot Series 3 is the empty seed dataset for the Jetson-only heavy model series. Series 3.x changes the learning target from steering-only control to joint steering and throttle control.

The dataset is intentionally empty until new smooth human-driving captures and controlled CARLA samples are collected. Series 3 should not be seeded with autonomous/model-predicted labels from the old Series 2.x runs.

Project code and documentation are maintained in the GitHub repo:

Resource	Link
GitHub repository	https://github.com/RamCodesBetter/SidewalkPilot
Hugging Face dataset	https://huggingface.co/datasets/ram-shreyas-naik-sabavat/SidewalkPilot_v3
Hugging Face model namespace	https://huggingface.co/ram-shreyas-naik-sabavat

Dataset Contents

File or folder	What it contains
sidewalkpilot_dataset/	Empty image folder reserved for Series 3 captures
steering_corrections.json	Empty list-style correction file for merged Series 3 labels
sidewalkpilot_trainer.py	Series 3 training, ONNX export, and TensorRT build script

Current Size

Item	Count
JPG images	0
Steering/throttle label entries	0
Label sources	0
Steering range	0 to 180 degrees
Throttle range	0.00 to 1.00

Label Format

steering_corrections.json starts as an empty JSON list:

[]

During Series 3 data collection, raw photo-run label files are saved separately as dict-style JSON files named like 2026_05_20_run_1.json. Those files map each captured image to the final steering and throttle commands at that frame.

Field	Type	Meaning
image filename key	string	Captured image filename
steering	number	Final steering servo angle in degrees
throttle	number	Final forward motor command

Example entry:

{
  "photo_20260520_123456.jpg": {
    "steering": 92,
    "throttle": 0.37
  }
}

When run data is promoted into steering_corrections.json, it may be converted into the trainer's list-style correction format.

Steering Label Meaning

The steering label is a servo angle in degrees.

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

Throttle Label Meaning

The throttle label is the final 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.

Planned Data Sources

Source	Status	Purpose
Smooth human driving	Planned	Main real-world imitation-learning labels
CARLA 50k with shadows	Planned	Controlled synthetic shadow and route coverage
Data augmentation	Planned	Training-time robustness, not stored as fake raw labels

Old Series 2.x autonomous/model-predicted labels are not valid Series 3 training labels.

Basic Loading Example

from pathlib import Path
import json

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

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

    print(image_path, steering_degrees, throttle)

Training Use

The labels are intended for the SidewalkPilot Series 3 trainer. The training target is:

image -> [steering, throttle]

Labels are stored in physical units. The trainer normalizes steering and throttle internally to the tanh range before loss calculation.

Series 3 defaults to 320x180 model input size. Throttle is required for every Series 3 training label; samples without throttle are skipped as bad labels.

Typical local training flow:

python3 sidewalkpilot_trainer.py \
  --roots sidewalkpilot_dataset \
  --corrections steering_corrections.json \
  --model-version 3.0 \
  --export-onnx

Jetson TensorRT INT8 build flow:

python3 sidewalkpilot_trainer.py \
  --roots sidewalkpilot_dataset \
  --corrections steering_corrections.json \
  --model-version 3.0 \
  --export-onnx \
  --build-tensorrt \
  --trt-precision int8

The trainer outputs normalized control vectors:

control_norm[0] = steering tanh output, -1.0 to 1.0
control_norm[1] = throttle tanh output, -1.0 to 1.0

Exact training commands may differ depending on CARLA data, source weighting, shadow augmentation, ONNX export, TensorRT conversion, and INT8 calibration.

Augmentation Preview

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

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

Evaluation Use

Series 3 evaluation should compare joint steering/throttle prediction quality. Common metrics should include:

Metric	Meaning
Steering MAE	Mean absolute steering error in degrees
Throttle MAE	Mean absolute throttle-command error
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

Intended Scope

This dataset supports the Series 3 Jetson-only research direction for SidewalkPilot. It is meant for heavy custom CNN regression models deployed through ONNX/TensorRT with INT8 optimization when calibration data is available.

The dataset should be updated only with synchronized image, steering, and throttle labels from smooth human driving or controlled CARLA collection.