TrajTok: Technical Report for 2025 Waymo Open Sim Agents Challenge

2025-06-23

AI Summary

To address insufficient trajectory coverage, poor symmetry, and weak robustness in behavioral generation models due to inadequate discrete trajectory modeling, this paper proposes TrajTok—a trajectory-space-structure-aware discretizer. TrajTok employs a hybrid discretization scheme integrating data-driven learning with physics-informed constraints and introduces, for the first time, a space-aware label smoothing loss to enhance training stability while preserving physical plausibility. Coupled with an improved cross-entropy loss and the SMART behavioral generation architecture, our method achieves a state-of-the-art realism score of 0.7852 on the 2025 Waymo Open Sim Agents Challenge. Key contributions include: (1) the first discrete tokenization mechanism explicitly encoding geometric trajectory structure; (2) a novel label smoothing strategy that jointly ensures spatial continuity and classification stability; and (3) end-to-end improvements in coverage, symmetry, and generalization robustness of behavioral generation.

Technology Category

• Humans and AI: Human-Aware Planning and Behavior Prediction
• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Multiagent Systems: Adversarial Agents

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

In this technical report, we introduce TrajTok, a trajectory tokenizer for discrete next-token-prediction based behavior generation models, which combines data-driven and rule-based methods with better coverage, symmetry and robustness, along with a spatial-aware label smoothing method for cross-entropy loss. We adopt the tokenizer and loss for the SMART model and reach a superior performance with realism score of 0.7852 on the Waymo Open Sim Agents Challenge 2025. We will open-source the code in the future.

Problem

• Develop trajectory tokenizer for behavior generation models
• Combine data-driven and rule-based methods for robustness
• Improve realism in Waymo Open Sim Agents Challenge

Innovation

• Combines data-driven and rule-based methods
• Uses spatial-aware label smoothing
• Adopts tokenizer for SMART model