ram-shreyas-naik-sabavat/SidewalkPilot_carla

SidewalkPilot CARLA Synthetic Dataset

CARLA-simulator-generated steering + throttle frames used to assist the SidewalkPilot Series 1/2 models — blended with real RC-car photos and down-weighted vs real. This is synthetic data rendered in the CARLA driving simulator, not real field capture.

Series 3 does NOT use this dataset — the Series 3 line is trained on real RC-car photos only. This CARLA set is kept for the CARLA-assisted Series 1/2 history and for optional future sim2real experiments.

Resource	Link
GitHub repository	https://github.com/RamCodesBetter/SidewalkPilot
Hugging Face dataset	https://huggingface.co/datasets/ram-shreyas-naik-sabavat/SidewalkPilot_carla
Real datasets	SidewalkPilot_v1_and_v2 (real S1/2) · SidewalkPilot_v3_and_v4 (shared real S3/4)

How this data was generated

The frames were rendered in the CARLA autonomous-driving simulator. A vehicle was driven along road/lane routes by an expert path-following controller while a front-facing camera logged each frame together with the control the expert applied. Every frame therefore pairs a rendered image with a clean expert steering + throttle label plus the controller's tracking state — that's what makes it usable for imitation learning (image → control).

The per-frame telemetry (speed, applied_steer, lateral_error, heading_error) is the fingerprint of that setup: a controller tracking a reference path, logging how much steering/throttle it applied and how far off the path it was (cross-track + heading error). Coverage spanned multiple CARLA towns and weather presets — the source folders were named dataset_carla_steering_town03_clear, ..._town04_cloudy, ..._town05_wet, etc. — giving varied roads, lighting, and surface conditions the small early real datasets lacked.

Project-specifics (CARLA version; exact town/weather split; capture resolution/fps; whether the expert was CARLA's built-in autopilot or a custom pure-pursuit/Stanley controller) belong to the SidewalkPilot generation setup. That generator is no longer in the repo (the old generate_synthetic_sidewalks helper was retired), so this set is preserved as the archived output.

Dataset Contents

File or folder	What it contains
sidewalkpilot_carla_dataset.tar	the full dataset as one tar — the 50,000 PNG frames + labels.json live inside (extract to reconstruct an images/ folder). This is how the images are hosted (no loose images/ folder on HF).
labels.json	per-frame labels (image, steering, throttle, + sim telemetry) — a loose copy alongside the tar for quick inspection
telemetry.json	extra per-frame simulator telemetry

Current Size

Item	Count
PNG images	50,000
Label entries	50,000
Steering range	0 to 180 degrees (logical; 90 = straight)
Throttle range	0.00 to 1.00
Source	CARLA simulator (synthetic)

Label Format

labels.json is a JSON list; each entry maps one frame in images/ to its controls plus sim telemetry:

Field	Type	Meaning
image	string	frame filename inside images/ (e.g. 000000.png)
steering	number	logical steering servo angle, 0–180
throttle	number	forward motor command, 0.00–1.00
speed	number	simulator speed (m/s)
applied_steer	number	normalized steering actually applied by the expert controller
lateral_error	number	cross-track error vs the reference path
heading_error	number	heading error vs the reference path

The SidewalkPilot trainer only reads image / steering / throttle; the rest is kept for analysis.

Example entry:

{
  "image": "000000.png",
  "steering": 90.477828,
  "throttle": 0.72,
  "speed": 0.49,
  "applied_steer": 0.00053,
  "lateral_error": 5.2e-08,
  "heading_error": -2.0e-07
}

How it assisted the Series 1/2 models

Early Series 1/2 real datasets were small (a few thousand hand-labeled field photos) and thin on turns, shadows, and route variety. CARLA filled those gaps:

• Volume + diversity: 50k synthetic frames across towns/weather added far more steering angles and lighting conditions than the real set alone.
• Blended, not dominant: the trainer tags any root whose name contains carla/synthetic/sim/dataset_l2 as source="carla" and down-weights it (--carla-sample-weight 0.6) vs real (2.0) — real data stays the anchor, CARLA is a supplement.
• Sim2real via domain randomization: CARLA frames get heavy augmentation (contrast, noise, blur, tree/edge shadows, texture — --carla-domain-randomize-probability 0.70) to bridge the render-vs-real gap so the model doesn't overfit the "clean sim look."
• Documented in the model cards: v1.0 = "initial mixed sidewalk/CARLA set"; v2.1 = "CARLA + real + corrections"; v2.2 = stronger shadow and CARLA/domain-randomization settings. It gave the baseline models turn + shadow coverage before enough real field data existed.

Series 3 dropped it — by then I had collected 50k+ real sidewalk photos, and Series 3 learns real-world steering+throttle directly (real-only).

How It's Used In Training

Trainer code is maintained in GitHub, not duplicated in this dataset repository. From a SidewalkPilot GitHub checkout, blend this data with a real dataset by listing both extracted roots in --roots, e.g.:

python3 sidewalkpilot_trainer.py --roots <real_dataset> carla_dataset --model-version <ver>

Series 1/2 models were trained this way (real + CARLA). Series 3 omits it.

Intended Scope

Synthetic sim2real assist data — not real field data, and not a standalone training set. Use it blended with, and down-weighted vs, real captures. Do not present CARLA predictions as real-world field performance.