DREAM: Defensive Risk-Aware Enhanced Maneuver Planning for Autonomous Vehicles in Heterogeneous Traffic

DREAM is an occlusion-aware, safety-critical planning framework that couples dynamic PDE risk transmission with maneuver-level decision-making and MPC-CBF control, explicitly capturing heavy-vehicle asymmetry and blind-zone hazards in heterogeneous traffic.

This repository is built upon the Risk Field Modeling Comparative Study and DRIFT: Dynamic Risk Inference via Field Transmission for Human-like Autonomous Driving.

proposed framework:

🚀 Quick Start

Step 1: Install required packages; Run Visualization Simulation based on the BEV dataset trajectories

cd src
pip install -r requirements.txt
python drift_dataset_visualization.py

Output:

• Frames saved to figsave_DRIFT_dataset/

Step 2: Create Video Animation

# You may change the file name every epoch you run the simulation and then save the video
python video_generation.py

Step 3: Analyze Risk Data

python risk_analysis_utils.py figsave_risk_viz/risk_at_ego.npy

Output:

• risk_timeline.png - Risk over time with threshold lines
• risk_histogram.png - Distribution of risk levels
• risk_analysis.png - Comprehensive multi-panel analysis
• risk_events.csv - High-risk events exported to CSV

---

🎨 Customizing Visualization

Option A: Use Presets (Easiest)

Edit emergency_test_with_risk_viz.py, add after imports:

from risk_viz_config import RiskVizConfig as viz_cfg

# Choose a preset
viz_cfg.preset_subtle()      # Low-contrast, clean
viz_cfg.preset_dramatic()    # High-contrast, emphasizes risk
viz_cfg.preset_scientific()  # Publication-ready with colorbar
viz_cfg.preset_highcontrast() # For presentations

# Then replace hardcoded values
RISK_ALPHA = viz_cfg.RISK_ALPHA
RISK_CMAP = viz_cfg.RISK_CMAP
# ... etc

Option B: Manual Tuning

Edit these variables in emergency_test_with_risk_viz.py (around line 140):

RISK_ALPHA = 0.4         # Transparency (0.0-1.0)
RISK_CMAP = 'hot'        # Colormap: 'hot', 'YlOrRd', 'plasma', 'inferno'
RISK_LEVELS = 15         # Number of contour levels
RISK_VMAX = 3.0          # Max risk value for color scale
SHOW_CONTOUR = True      # Show contour lines?
SHOW_HEATMAP = True      # Show filled heatmap?

Colormap Options:

• 'hot' - Black → Red → Yellow → White (classic heat)
• 'YlOrRd' - Yellow → Orange → Red (warning colors)
• 'Reds' - White → Red (simple gradient)
• 'plasma' - Purple → Pink → Yellow (perceptually uniform)
• 'inferno' - Black → Purple → Orange → Yellow
• 'RdYlGn_r' - Red → Yellow → Green (reversed)

---

📊 Understanding the Risk Field

DRIFT Risk Sources

The risk field combines three sources:

1. Vehicle-Induced Risk (Anisotropic Gaussian kernels)

   • Higher in front of moving vehicles (direction of travel)
   • Decays with distance
   • Intensity scales with relative velocity

2. Occlusion-Induced Risk (Shadow regions)

   • Elevated behind large vehicles (trucks, trailers)
   • Represents hidden hazards in sensor blind spots
   • Propagates based on uncertainty

3. Merge Pressure (Topological conflicts)

   • High in lane-change zones
   • Elevated where lanes converge
   • Captures structural road geometry risks

Risk Propagation Dynamics

• Advection: Risk flows with traffic
• Diffusion: Risk spreads spatially (uncertainty)
• Telegraph term: Finite propagation speed (wave-like)

---

🔍 Interpreting Results

High Risk Scenarios

You should see elevated risk (red zones) in:

1. Dense traffic - Multiple vehicles close together
2. Behind large vehicles - Occlusion shadows
3. Lane change maneuvers - Merge pressure zones
4. Emergency braking - Sudden deceleration events

Analysis Metrics

• Mean Risk: Average exposure level
• Peak Risk: Maximum risk encountered
• Time in High Risk: Duration above threshold
• Risk Events: Discrete high-risk episodes

---

📐 Technical Details

Simulation Parameters

• Grid: From config.py (default: 400m × 60m, 1m resolution)
• PDE Substeps: 3 (for numerical stability)
• Timestep: 0.1s (matches IDEAM)
• Horizon: 400 timesteps (40 seconds)

Performance

• Frame generation: ~2-3 seconds per frame (depends on grid size)
• Full simulation: ~15-20 minutes for 400 frames

---

📈 Example Workflow

1) Emergency highway scenario (synthetic)

python emergency_test_prideam.py \
  --integration-mode conservative \
  --steps 120 \
  --scenario-file file_save/120_100 \
  --save-dir outputs/emergency_run01 \
  --save-dpi 300 \
  --save-frames true

2) Uncertainty merger scenario (synthetic)

python uncertainty_merger_DREAM.py \
  --integration-mode conservative \
  --steps 120 \
  --save-dir outputs/uncertainty_merger_run01 \
  --save-dpi 300 \
  --save-frames true

3) Dataset benchmark (rounD/inD replay)

python dream_dataset_benchmark.py \
  --dataset-dir data/rounD \
  --recording-id 01 \
  --ego-track-id 254 \
  --save-dir outputs/dataset_benchmark_run01 \
  --steps 120 \
  --integration-mode conservative \
  --save-frames true \
  --frame-dpi 150

Optional: frame sequence to MP4

python video_generation.py \
  --image-folder outputs/dataset_benchmark_run01 \
  --video-name outputs/dataset_benchmark_run01/benchmark.mp4 \
  --fps 20

Baseline parallel simulations (simulations in high-density traffic):

# All 5 arms (default)
python uncertainty_test_DREAM.py

# Specific arms
python uncertainty_test_DREAM.py --models DREAM IDEAM
python uncertainty_test_DREAM.py --models OA-CMPC IDEAM
python uncertainty_test_DREAM.py --models DREAM ADA APF OA-CMPC IDEAM

# Override run mode too
python uncertainty_test_DREAM.py --models DREAM IDEAM --mode batch

Demonstrations:

demonstration of LC for emergency vehicle with safety-critical considerations (IDEAM-based planning).

Compared with the baseline planner, DREAM enables the ego stay away from the agent group ahead and find the appropriate spaces with no agents around, where the risk score is minimal. However, the progress was sacrificed.

The baseline planner forced the ego to perform very aggressive and dangerous overtaking (LC to the left) and nearly collide with the rear of the truck-trailer ahead. Instead, DREAM shows a more conservative planning that better aware of the risk from the truck-trailer and the uncertainty.

We compare the trajectories of (1): the ground truth ego trajectories from BEV datasets; (2): the baseline planner trajectories; (3): the DREAM planner trajectories. The results show that baseline planner is over aggressive as near collision with the truck rear, and our planner is more conservative but sacrifice the progress. (The selected scenario include the occlusion-aware planning from the truck-trailer that may block the visibility of the ego)

Acknowledgement:

The BEV dataset visualizations:

drone-dataset-tools

The baseline IDEAM (MPC-CBF) planner:

Corresponding paper:

@article{shu2025agile,
  title={Agile Decision-Making and Safety-Critical Motion Planning for Emergency Autonomous Vehicles},
  author={Shu, Yiming and Zhou, Jingyuan and Zhang, Fu},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2025},
  publisher={IEEE}
}

(Referenced coding package: IDEAM)

The baseline Artificial Potential Field (APF) modeling:

Corresponding paper:

@article{gao2025trajectory,
  title={Trajectory Planning Algorithm Considering Obstacle Risk in Dynamic Traffic Scenarios},
  author={Gao, Aiyun and Zhang, Wei and Fu, Zhumu and Tao, Fazhan},
  journal={IEEE Transactions on Vehicular Technology},
  year={2025},
  publisher={IEEE}
}

(Referenced coding package: Artificial-Potential-Field)

The baseline Gaussian Velocity Field (GVF) modeling:

@article{zhang2021spatiotemporal,
  title={Spatiotemporal learning of multivehicle interaction patterns in lane-change scenarios},
  author={Zhang, Chengyuan and Zhu, Jiacheng and Wang, Wenshuo and Xi, Junqiang},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={23},
  number={7},
  pages={6446--6459},
  year={2021},
  publisher={IEEE}
}

(Referenced coding package: Gaussian Velocity Field (GVF))

The baseline Asymmetric Driving Aggressiveness (ADA) modeling:

@article{hu2025socially,
  title={Socially Game-Theoretic Lane-Change for Autonomous Heavy Vehicle based on Asymmetric Driving Aggressiveness},
  author={Hu, Wen and Deng, Zejian and Yang, Yanding and Zhang, Pingyi and Cao, Kai and Chu, Duanfeng and Zhang, Bangji and Cao, Dongpu},
  journal={IEEE Transactions on Vehicular Technology},
  year={2025},
  publisher={IEEE}
}
@article{deng2024eliminating,
  title={Eliminating uncertainty of driver’s social preferences for lane change decision-making in realistic simulation environment},
  author={Deng, Zejian and Hu, Wen and Sun, Chen and Chu, Duanfeng and Huang, Tao and Li, Wenbo and Yu, Chao and Pirani, Mohammad and Cao, Dongpu and Khajepour, Amir},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2024},
  publisher={IEEE}
}

The baseline occlusion-aware contingency MPC (OA-CMPC):

@article{zhengocclusion,
  title={Occlusion-Aware Contingency Safety-Critical Planning for Autonomous Driving},
  author={Zheng, Lei and Yang, Rui and Zheng, Minzhe and Peng, Zengqi and Wang, Michael Yu and Ma, Jun},
  journal={IEEE transactions on cybernetics},
  pages={1-14},
  year={2026},
  publisher={IEEE}
}

Citation (preprint):

@article{zianDREAM,
  title={DREAM: Defensive Risk-Aware Enhanced Maneuver Planning for Autonomous Vehicles in Heterogeneous Traffic},
  author={Zian, Wang and Yiming, Shu and Zejian, Deng and Guoshun, Cai and Jiahui, Xu and Jiwei, Tang and Dongpu, Cao and Sun, Chen},
  year={2026},
  doi={https://doi.org/10.2139/ssrn.6500569},
  journal={SSRN Preprint}
}