Teaching AVs the Language of Human Driving Behavior with Trajeglish

• The NVIDIA Research team has developed an approach called Trajeglish to simulate realistic multi-vehicle driving scenarios by tokenizing motion.
• Trajeglish tokenizes a given scenario, enabling the model to predict likely trajectories based on scene context.
• By analyzing the distribution of all tokenized scenarios, Trajeglish models the next actions in a timestep, taking into account intra-timestep dependence.
• In the Waymo Sim Agents Challenge, Trajeglish outperformed 16 other models, improving the overall realism of scenarios by 3.3% and topping the interaction component by 9.9%.
• The approach borrows from language modeling, which deals with complexities in human language, to address the nuanced nature of human driving behavior.