Driving the intelligent era with AV2.0
Plus’s AV2.0 approach moves away from a traditional code-based architecture to a data-driven framework that uses minimal code and leverages sophisticated neural network models.
Our model-centric strategy enables us to generalize, using vast datasets encapsulating a wide range of human driving experiences and knowledge. These large AI models allow us to achieve the complex levels of dynamic reasoning necessary for Level 4 autonomy and what powers our robust and adaptive autonomous driving systems.
Driving Intelligence
Our driving intelligence models are built using Gen AI, open foundation models, and proprietary data to understand the physical world and make real-time predictions.
Our model implicitly learns the physics of motion, concepts of gait, depth and reflection to then predict the future behaviors of vehicles and humans.
Efficient Training
.gif&w=3840&q=75)
Our innovative auto-labeling and model distillation techniques enable more efficient training of AI models by reducing time and cost of learning.
We use multi-view camera auto-labeling algorithms to produce 3D reconstructed ground-truth data to fine-tune our AI models.
Power Efficiency
Our state-of-the-art network design allows us to provide industry-leading high-performance DNN models that can run across different automotive grade SoCs.
Our 200m+ detection range model can scale from ~20 TOPS (1 camera at 10 fps) to ~220 TOPS (11 cameras at 30 fps).
Structure of Our Approach
1. Gen-AI model for Decision Making
Our system employs advanced models to generate the best candidate driving decisions, ensuring each decision is optimized for the given circumstances.
2. Perception Layer Abstraction
Our system generates a structured abstraction of the real-world environment that is both coherent and interpretable, enabling formal safety checks.
3. Safety Envelope
Our rule-based safety envelope allows tailored behavior for different regions and provides strong guarantees for edge-cases.
Benefits of Our Approach
Interpretability
The defined perception layer allows for clearer explanations of the autonomous system’s decisions, making it easier for engineers and stakeholders to understand and trust the technology.
Tractability
Enhanced interpretability makes it easier to trace and rectify errors or exceptions, particularly in identifying and addressing corner cases, thereby continually improving the system.
Enhanced Safety
Driving is inherently safety-critical. Our robust safety envelope, inspired by foundational safety principles, such as Asimov’s rules for robotics, ensures our system adheres to strict safety protocols.
Scalability
The flexibility of our approach allows for easy adaptation across different platforms, applications, and customer needs. This scalability is particularly valuable as it enables customers to integrate our technology according to their specific requirements, whether that involves using our complete system or incorporating individual components like the perception model.