The quest for safe and efficient autonomous driving hinges on rapid, high-fidelity trajectory generation. Existing diffusion-based methods, while powerful, are bottlenecked by significant inference latency stemming from their iterative nature. This limitation poses a critical challenge for real-time deployment.
From Iterative Diffusion to Single-Step Synthesis
The researchers introduce MISTY (Mixer-based Inference for Single-step Trajectory-drifting Yield), a novel generative motion planner designed for high-throughput, single-step inference. By integrating a vectorized Sub-Graph encoder for context, a Variational Autoencoder to distill expert trajectories into a compact latent space, and an ultra-lightweight MLP-Mixer decoder, MISTY circumvents the quadratic complexity of attention mechanisms. This architectural shift is key to its dramatic speed improvements.
Latent-Space Drifting for Proactive Maneuvers
A core innovation is the latent-space drifting loss. This mechanism elegantly shifts the complex distribution evolution from inference time to training. By formulating explicit attractive and repulsive forces within the latent space, the MISTY motion planner is empowered to generate not just safe but also proactive maneuvers, such as active overtaking, which are often underrepresented in raw expert data. This capability is crucial for developing truly intelligent and adaptable autonomous systems.
Industry-Leading Performance and Throughput
Evaluated on the demanding nuPlan benchmark, MISTY achieves state-of-the-art results on the challenging Test14-hard split, securing comprehensive scores of 80.32 in non-reactive and 82.21 in reactive settings. Crucially, it operates at over 99 FPS with an end-to-end latency of just 10.1 ms. This represents an order-of-magnitude speedup compared to iterative diffusion planners, making the MISTY motion planner a compelling solution for real-world autonomous driving applications where split-second decisions are paramount.
