DA-WAM: Decision-Aligned Future Latents for Driving World Models

Noosaga Placements

• Roboticssubfield95%
  The paper focuses on autonomous driving, a core robotics application, and is categorized under cs.RO. It addresses trajectory planning and world modeling for driving, which are central to robotics.
  Anticipating how scenes evolve under ego actions is fundamental to safe autonomous drivingarXiv:2608.19085v1 [cs.RO]
• Learning-Based Roboticsframework90%
  The paper presents a learning-based approach for autonomous driving, using neural networks for prediction and planning, which aligns with the Learning-Based Robotics framework.
  DA-WAM, a framework that unifies predictive representation learning, action-conditioned future modeling, and trajectory scoringonline encoder, predictor, and scorer
• Model-Based Learning and Planningframework85%
  The method uses a world model to predict future states and uses these predictions for planning, which is a model-based learning and planning approach.
  world models for decision-makingaction-conditioned future modeling
• Robot Learningsubfield80%
  The method involves learning-based approaches for future prediction and trajectory scoring, including representation learning and adaptation, which are key aspects of robot learning.
  DA-WAM maintains predictive supervision throughout planner optimization via an online encoder and a stable momentum targetallowing future representations to co-evolve with the driving task
• Deep Learningframework80%
  The framework employs deep learning techniques, including transformers and LoRA, for representation learning and prediction.
  V-JEPA 2.1 backboneLoRA-adapted online encoder
• Machine Learningsubfield70%
  The paper uses self-supervised learning (JEPA), representation learning, and supervised training for scoring, which are machine learning techniques.
  JEPA-driven predictive representation adaptationLoRA-adapted V-JEPA 2.1 online encoder
• Self-Supervised Learningframework80%
  The predictive representation learning is based on self-supervised learning principles, as seen in JEPA, which is a self-supervised learning framework.
  JEPA-driven predictive representation adaptationpredictive supervision throughout planner optimization
• Deep Learningsubfield70%
  The framework relies on deep neural networks, including transformers and LoRA adaptation, for encoding, prediction, and scoring.
  V-JEPA 2.1 backbonescoring transformer cross-attends

Abstract

Anticipating how scenes evolve under ego actions is fundamental to safe autonomous driving, yet the full potential of world models for decision-making remains unrealized. The critical challenge lies in ensuring that future modeling is not merely predictive, but decision-informative: the predicted future must directly shape which trajectory is selected. Existing approaches decouple future representation learning from planning optimization, or share predicted states across trajectory candidates, thereby diluting the action-specific consequences that ought to guide selection. To bridge this gap, we propose DA-WAM, a framework that unifies predictive representation learning, action-conditioned future modeling, and trajectory scoring under a single decision-making objective. DA-WAM maintains predictive supervision throughout planner optimization via an online encoder and a stable momentum target, allowing future representations to co-evolve with the driving task. An action-conditioned predictor generates a distinct future latent state per trajectory candidate, which is then evaluated by a future-latent-conditioned factorized scorer. For the expert-matched trajectory, the predicted future latent is supervised by the observed future representation, while safety-critical hard negatives provide additional supervision near planning boundaries. Extensive experiments on NAVSIM-v1 and NAVSIM-v2 demonstrate state-of-the-art performance, while ablations and diagnostic analyses validate the key components.

Paper Context