E2E Platform (Autonomous Driving)

Next-generation autonomous driving platform based on a massive single neural network AI

Next-generation autonomous driving platform based on a massive single neural network AI

Moving beyond separated rules (Rule-base), it achieves perfect autonomous driving flowing from perception to driving planning through a single neural network.

e2e Platform architecture Image
What is the E2E Platform (Autonomous Driving)?

The ultimate destination for next-generation autonomous driving

Traditional autonomous driving methods were 'Lego blocks' with all rules coded by hand.

The E2E Platform (Autonomous Driving) operates as an organism where AI multi-dimensionally perceives the world, makes its own judgments, and moves the vehicle.

1

Integrated Spatial Representation

  • check Real-time conversion of multiple camera and sensor data into 3D top-view (BEV) space
  • check Mapless driving based on real-time sensor information
2

Feature-Level Data Flow

  • check Maintains data flow in a high-dimensional feature vector state without structured data processing
3

Plan-Centric Global Optimization

  • check Complete model training based on the ultimate goal of 'safe driving planning'
  • check Learning psychological interaction with other vehicles
4

Explainable Auxiliary Supervision

  • check Simultaneous execution of auxiliary tasks such as 3D object detection and occupancy grid prediction alongside driving path generation
Prototype 6-Camera
Prototype Forecast
Prototype

Aiming for an integrated inference (engine) for E2E.

Graph_3

4-in-1

A platform that performs perception + mapping + prediction + planning simultaneously in a single model (1 Encoder and P3 Decoder)

neurology

A Single Neural Network

6-camera 3D perception and mapping, vehicle trajectory prediction, and driving planning inferred by a single neural network

Timer

Precise Prediction

Simultaneous prediction of surrounding vehicle trajectories (up to 6 seconds) and ego vehicle driving plans (3 seconds)

School

AI Learning-Based

Predicted results match the ground truth even in left-turn situations - implemented through a learning method, not rule-based

NFLUX (End-to-End)

Integrated Inference (Engine) Prototype

Perception, mapping, prediction, and planning are processed by a single neural network. It automatically adapts to complex situations based on learning and saves real-time computing resources.

항목
NFLUX E2E 방식
업계 일반 방식
Number of Models
check_circle 1
3
Processing Method
check_circle Simultaneous
Sequential
Manual Rule Adjustment
check_circle Not Required
Required
Adaptation to New Situations
check_circle Auto-learning
Re-coding
※ 상기 영상은 초당 1frame 으로 편집되어있어, 실제 영상과 다소 차이가 있을 수 있습니다.
Roadmap

2026 ~ 2030 Mid-to-Long Term Business Roadmap

  • Phase 1: The Beginning of Next-Gen Autonomous Driving Platforms

    1. E2E integrated inference engine prototype (1 encoder + 3 decoders)
    2. 6-camera 3D perception & BEV mapping
    3. Rule-based, learning-based, integrated spatial representation

  • Phase 2: Real-Road Data and Performance Expansion

    1. Real-road data fine-tuning
    2. 9x reduction in collision risk
    3. API international standardization research

  • Phase 3: Preparation for Commercial Advancement

    1. Model optimization for edge computers
    2. Acquisition of temporary driving permit
    3. Seed investment
    4. Domain scenario learning

  • Phase 4: Global Leap

    1. Evolution of VFM (Vision Foundation Model)
    2. Global data connection
    3. Contracts with global OEMs