ACTIVE Research

DriveMind-Prototype

Intelligent driving decision framework combining LLM reasoning with physics models for 4500x faster decisions

// DESCRIPTION

DriveMind bridges the gap between high-level LLM reasoning and real-time driving decisions through a novel hybrid architecture.

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    Scene Understanding                       │
│              (Camera, LiDAR, HD Map, Traffic)               │
├─────────────────────────────────────────────────────────────┤
│   ┌─────────────────┐      ┌─────────────────┐             │
│   │ LLM Reasoning   │ ←──→ │ Physics Model   │             │
│   │ (Strategy)      │      │ (Risk Field)    │             │
│   └─────────────────┘      └─────────────────┘             │
│            ↓                        ↓                       │
│   ┌─────────────────────────────────────────────┐          │
│   │         Decision Fusion Layer               │          │
│   │    (LLM intent + Physics constraints)       │          │
│   └─────────────────────────────────────────────┘          │
│                        ↓                                    │
│              Motion Planning & Control                      │
└─────────────────────────────────────────────────────────────┘

Key Innovation

  • LLM for Strategy: High-level decision making (lane change, overtake)
  • Physics for Safety: Risk field ensures collision avoidance
  • Hybrid Fusion: LLM intent constrained by physical feasibility
  • 4500x Speedup: Pre-computed LLM decisions + real-time physics

Dataset Support

  • highD: German highway recordings
  • inD: Intersection scenarios
  • rounD: Roundabout maneuvers
  • CARLA: Simulation integration

Performance

MetricValue
Decision Latency<1ms
Safety Violation Rate0.02%
Scenario Coverage95%+

// HIGHLIGHTS

  • Nature Communications Risk Field integration
  • Multi-dataset evaluation framework
  • Academic paper in preparation
  • CC BY-NC-SA 4.0 License

TECH_STACK

Python PyTorch Ollama/OpenAI CARLA highD/inD/rounD

PROJECT_INFO

started: 2024-06-01
status: ACTIVE
type: Research