Daniel Ortiz Valencia
Portfolio

An incremental perception and control pipeline for autonomous navigation in the Formula Student Driverless Simulator (FSDS). A reactive LiDAR-only baseline isolates cone positions at 20Hz with O(n) sequential clustering and a binary density heuristic for steering, then a forward-facing camera processed by a YOLOv5 detector with FSOCO weights is fused with LiDAR depth. Pinhole ground-plane projection and nearest-neighbor matching align the two sensors to a 0.06m average match distance with near-zero systematic bias. A three-tier hybrid color verification scheme (stripe brightness, BGR channel dominance, HSV hue fallback) reaches roughly 98% accuracy at close range, and a color-aware proportional controller tracks the geometric centerline between the blue and yellow boundaries, eliminating the straight-line oscillation of the baseline while keeping a graceful fallback to the density heuristic under camera failure.

A two-phase system that detects and suppresses YOLO color misclassifications before they reach the path planner. Phase 1 builds a ground-truth dataset: a ROS2 bridge and four-stage spatial matching pipeline label 129,025 cone detections from the Formula Student Driverless Simulator, exposing a 2.52% baseline false-positive rate with orange cones failing at 24.4% (9.7x the mean). Phase 2 trains a two-model XGBoost Safety Gate, framed as supervised anomaly detection, on 17 features including five engineered context features (neighbor agreement, lateral outlier, relative size, corner flag, and a corner-by-prior interaction). On the held-out test set the system reaches F1 = 0.906, PR-AUC = 0.976, and ROC-AUC = 0.999, cutting incorrect color labels reaching the steering controller from 2.73% to roughly 0.24%, a 91% reduction in the failure mode that causes wrong lane assignment, at under 0.3ms per frame on a Raspberry Pi 4.

A study of replacing a hand-coded reactive steering controller with a learned policy in the Formula Student Driverless Simulator. The work delivers a reproducible RL environment with a 41-dimensional cone observation, a single steering action, and a three-term reward, with the policy trained against ground-truth cones and warm-started by behavior cloning to isolate control from perception. Soft Actor-Critic was evaluated across six configurations and every run ended in collapse, traced to three structural properties of the algorithm rather than tuning: replay-buffer poisoning, automatic-entropy collapse, and catastrophic forgetting after reward changes. Proximal Policy Optimization is then adopted, whose on-policy formulation, trust-region clipping, and fixed entropy coefficient remove each failure mode by design at the cost of more environment interactions. The result is a validated training infrastructure and a catalogue of failure modes with recognizable training-log signatures.