MISTY: Single-Step Motion Planner for Autonomous Driving
A team of researchers has introduced MISTY (Mixer-based Inference for Single-step Trajectory-drifting Yield), an advanced generative motion planner designed for autonomous vehicles that excels in closed-loop performance through pure single-step inference. In contrast to current diffusion-based planners that experience delays from iterative evaluations of neural functions, MISTY employs a vectorized Sub-Graph encoder for environmental context, a Variational Autoencoder that condenses expert trajectories into a streamlined 32-dimensional latent space, and a lightweight MLP-Mixer decoder that removes quadratic attention complexity. A significant advancement is the latent-space drifting loss, which transfers the evolution of complex distributions to the training stage by establishing clear attractive and repulsive forces, allowing the model to create innovative, proactive trajectories. This research is published in a paper on arXiv (2604.21489).
Key facts
- MISTY is a generative motion planner for autonomous driving.
- It achieves state-of-the-art closed-loop performance with single-step inference.
- Uses a vectorized Sub-Graph encoder for environment context.
- Employs a Variational Autoencoder to structure expert trajectories into a 32-dimensional latent manifold.
- Features an ultra-lightweight MLP-Mixer decoder to eliminate quadratic attention complexity.
- Introduces a latent-space drifting loss with explicit attractive and repulsive forces.
- The loss shifts complex distribution evolution to the training phase.
- Paper available on arXiv with ID 2604.21489.
Entities
Institutions
- arXiv