Autonomous Driving in TORCS — Behavioral Cloning with K-NN (KD-Tree)

Autonomous driver for TORCS built with Behavioral Cloning + K-Nearest Neighbors, accelerated via KD-Tree search and designed for robustness (fallback + stuck detection + OOD guards).

Repo: https://github.com/landigf/TORCS

Portfolio/educational project. No credentials or proprietary datasets are committed.

---

Demo

Cinematic run

---

Feature selection (data-driven)

We selected a compact, low-redundancy feature set using:

• Correlation matrix (to reduce redundant inputs)
• Mutual information (to keep the most informative features for control)

Track sensors used (subset of the 19 front sensors):

---

Core idea

1. Collect a dataset from manual driving (TORCS sensors + human actions).
2. Normalize features so Euclidean distance is meaningful and stable.
3. Build a KD-Tree model and serialize it to a .tree file.
4. Drive autonomously by retrieving nearest expert actions in real time, with safety fallback logic.

---

Key features

• Behavioral Cloning with K-NN: imitation learning via nearest-neighbor matching (state → action).
• KD-Tree search + serialization: faster than linear scan; models saved as .tree.
• Track segmentation: multiple KD-Trees by distanceFromStart for higher local coherence and lower latency.
• Weighted distance + dynamic k: stability/precision tradeoff depending on driving context.
• Robustness mechanisms: action caching, out-of-distribution (OOD) thresholding, latency guard, off-track recovery, and stuck detection with rule-based fallback.

---

Quickstart (4 steps)

TORCS SCR server typically listens on localhost:3001. Adjust host/port if needed.

Step 1 — Collect the dataset (manual driving logs)

javac -d classes src/scr/*.java src/scr/ai/SimpleGear.java src/scr/ai/DataLoggerDriver.java
cd classes
java -cp . scr.Client scr.ai.DataLoggerDriver localhost:3001 verbose:on

Remote example:

java -cp . scr.Client scr.ai.DataLoggerDriver host:172.19.196.17 port:3001 verbose:on

(Optional) Step 1.a — Clean/validate dataset

python .utility/dataset_clean.py

---

Step 2 — Build the KD-Tree model (.tree)

javac -d classes src/scr/ai/DataPoint.java src/scr/ai/KDTree.java src/scr/ai/DatasetBuilder.java
cd classes
java -cp . scr.ai.DatasetBuilder ../classes/dataset_union.csv knn.tree

---

Step 3 — Compile the K-NN driver

javac -d classes -cp classes src/scr/ai/ActionCache.java src/scr/ai/KNNDriver.java

---

Step 4 — Run autonomous driving

cd classes
java -cp . scr.Client scr.ai.KNNDriver localhost:3001 verbose:on

---

Build script (PowerShell)

.\.build_knn.ps1
.\.build_knn.ps1 -Config sensors
.\.build_knn.ps1 -Config all
.\.build_knn.ps1 -Config sensors -DatasetPath "my_dataset.csv" -ModelName "my_model.tree"

---

Repository structure (as-is)

• src/ — Java sources (drivers, KDTree, dataset builder, etc.)
• .utility/ — Python utilities (dataset cleaning, analysis)
• classes/ — compiled output (generated by build steps)
• .build_knn.ps1 — build helper script

---

Authors

• Gennaro Francesco Landi
• Maurizio Melillo
• Elettra Palmisano