maxim-igenbergs/dave2-gru

DAVE-2-GRU End-to-End Driving Model

DAVE-2 architecture extended with GRU (Gated Recurrent Unit) for temporal modeling, trained on the Udacity self-driving car simulator for the bachelor's thesis: Dual-Axis Testing of Visual Robustness and Topological Generalization in Vision-based End-to-End Driving Models.

Model Description

DAVE-2-GRU extends the original NVIDIA DAVE-2 architecture by adding a GRU layer to capture temporal dependencies across frames. This allows the model to leverage sequential information for smoother and more context-aware driving predictions.

Architecture

Input: RGB Image Sequence (N × 66 × 200 × 3)
    ↓
TimeDistributed:
  Conv2D(24, 5×5, stride=2) + ELU
  Conv2D(36, 5×5, stride=2) + ELU
  Conv2D(48, 5×5, stride=2) + ELU
  Conv2D(64, 3×3) + ELU
  Conv2D(64, 3×3) + ELU
  Flatten
    ↓
GRU(128)
    ↓
Dense(100) + ELU
Dense(50) + ELU
Dense(10) + ELU
    ↓
Output: [steering, throttle]

Checkpoints

Map	Checkpoint
GenRoads	genroads_20251215-174930/
Jungle	jungle_20251201-142321/

Files per Checkpoint

• best_model.h5: Keras model weights
• meta.json: Training configuration and hyperparameters
• history.csv: Training/validation metrics per epoch
• loss_curve.png: Visualization of training progress

Citation

@thesis{igenbergs2026dualaxis,
  title={Dual-Axis Testing of Visual Robustness and Topological Generalization in Vision-based End-to-End Driving Models},
  author={Igenbergs, Maxim},
  school={Technical University of Munich},
  year={2026},
  type={Bachelor's Thesis}
}

Related

• DAVE-2 Driving Model
• ViT Driving Model
• TCP Driving Model
• Training Data
• Evaluation Runs