GPT Trajectory Generation for Low-Latency Autonomous Vehicle Planning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing trajectory generation models for autonomous vehicles are complex, require significant tuning, and suffer from high latency and large model sizes, hindering efficient navigation.
Innovation Solution
Implementing a Generative Pre-trained Transformer (GPT)-based model for trajectory generation that utilizes vector maps and vectorized intent features, with an encoder-decoder transformer architecture to simplify architecture, reduce latency, and minimize tuning parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If complex trajectory generation models are used to provide multiple feasible trajectory options, then the quality and diversity of trajectory predictions are improved, but the model size and computational complexity increase significantly
Solution Approach 1:
The trajectory generation model is divided into an encoder component and a decoder component. The encoder processes input data (vehicle state, map information, sensor data) and generates latent representations, while the decoder generates multiple trajectory options from these representations. This segmentation allows each component to be optimized independently, reducing overall model complexity while maintaining the ability to generate diverse trajectory options.
Solution Approach 2:
A latent space or intermediate representation is introduced between the input data and the generated trajectories. The encoder maps input data to this intermediate latent space, which captures essential features without requiring the full complexity of the original data. The decoder then generates multiple trajectory options from this simplified intermediate representation, reducing computational burden while preserving trajectory quality.
2Measurement precision
If traditional trajectory generation models are used to ensure accurate navigation, then the prediction accuracy is improved, but the latency increases and real-time performance deteriorates
Solution Approach 1:
The encoder pre-processes and compresses input data into compact latent representations before trajectory generation. This preliminary action reduces the computational burden on the decoder, enabling faster trajectory generation while preserving the essential information needed for accurate predictions. The encoder acts as a preprocessing stage that prepares data in an optimized format for subsequent decoding.
Solution Approach 2:
The model transforms the input data into a different parameter space (latent space) with fewer dimensions and more compact representation. This parameter transformation reduces the computational complexity of trajectory generation while maintaining prediction accuracy. The latent representations capture the essential features in a more efficient format that requires less computational resources to process.
3Adaptability or versatility
If large model sizes are used to capture complex driving scenarios, then the coverage of edge cases is improved, but the memory requirements and processing overhead increase
Solution Approach 1:
The encoder extracts only the most relevant and essential features from the input data, separating them from redundant or less important information. This extraction process creates compact latent representations that capture the essential characteristics of driving scenarios without requiring large model sizes. The decoder then generates trajectories based on these extracted essential features, maintaining scenario coverage with reduced model complexity.
Data Source
AI summary
Disclosed are embodiments for facilitating generative artificial intelligence to generate multiple autonomous vehicle future trajectories. In some aspects, an embodiment includes receiving input data to a generative pre-trained transformer (GPT)-based trajectory generation model, wherein the input data comprises vector map representations, nearby actor history, and autonomous vehicle (AV) history of an AV; generating map tokens from the vector map representations and generating agent tokens from the nearby actor history and the AV history; inputting a concatenated set of the map tokens and the agent tokens into an encoder transformer of the GPT-based trajectory generation model; outputting, by the encoder transformer, an output embedding that is representative of a scene of the AV; and determining, by a decoder of the GPT-based trajectory generation model, a sequence of AV waypoint predictions for the AV based on the output embedding.