Gemini AI
Full Autonomy Trucks
Achieving Full Autonomy
Using a software-centric approach that incorporates self-learning and sensor fusion, Gemini AI runs vehicles autonomously, performs predictive maintenance, maximizes fuel efficiency, and meets the highest safety standards.
GDrive
Advanced AI technology for vehicle autonomy using a vision-centric, self-learning framework.
Self-Supervised
Learns from sensory data with minimal need for human supervision.
Scalable
Extracts relevant information and quickly adapts to new environments - enabling geographical coverage expansion.
Reliable
Enables better predictions with detailed hierarchical object-based representations.
Graph-based
Vision System
Uses prediction as a reward mechanism to transform high-dimensional spatio-temporal video data into more accurate graph structures.
GDesign
Reinforcement-learning based technology optimizes end-to-end system design and performance.
Data Fusion
Integrates data and physics to model and optimize complex systems.
High Performance
Provides scalable learning algorithms for optimized end-to-end designs.
Unmatched Efficiency
Unlocks new possibilities for efficient vehicle design and manufacturing.
Learning Framework
Learns from interactions with various simulators and experimental data by utilizing a reinforcement-learning based architecture.
Inside the AI
Graph-Structured Representation
Our technology utilizes hierarchical representations to provide autonomous robots with redundant models that improve their resiliency in navigating complex environments. It transforms videos into hierarchical graphs with multiple layers where each layer is connected to a coarser layer through a part-whole relationship.
This approach enables more efficient storage and faster learning by significantly reducing the data size and dimensions, thereby making our technology more scalable as compared to standard pixel-based representations.
VideoGraph enables the Gemini AI to:
- Learn from the data at scale
- Not require decisions from the programmer and explicit labels
- Continuously isolate and learn from the informative sensory data
Label-Free Segmentation
Our vision-based framework segments objects without the need for explicit labels from annotators.
This system employs basic principles such as time-continuity associated with motion, separation of objects based on their sound, and statistical co-occurrences of juxtaposed objects to segment raw videos into objects and their associated properties.
Agent-Based Optimization
GDesign’s high performance reinforcement-learning optimizer combines learning across GPUs and interactions with different simulators.
It enables large scale optimizations on top of digital twins, and extends the capability of Gemini in developing the most efficient designs.
Long-Term Prediction
Gemini AI predicts longer into the future by utilizing the combined information in the hierarchy of representations.
The improved prediction enhances motion planning to be more robust and trustworthy. This is important for planning in environments with many biological agents whose behaviors are not easy to predict.
Our AI learns by comparing the predictions and the observations and adjusting its segmentation and graph representations