New diffusion axis: from noise to quality
Instead of relying on a classic approach where the model removes Gaussian noise from the trajectory, the new QualiDrive model introduces an innovative mechanism - diffusion along the axis of driving quality. Instead of treating quality as an additional guiding signal, researchers decided to use it as the central axis of the trajectory generation process. The model builds so-called "quality ladders" for each scene, comparing and ordering known trajectories according to their actual driving quality - i.e., how safely, smoothly, and efficiently they were performed.
Each reverse step of the model is conditioned by the current scene, vehicle state, and desired level of quality. This allows for precise adaptation of the trajectory to specific conditions - e.g., in dense traffic or when changing lanes. As a result, the model not only generates a trajectory but develops it step by step from low to high quality, which has an impact on the final quality. Research shows that this approach is not just theoretical - it has a real impact on the results of tests.
Results on NAVSIM: quality outperforms average
Enlarged imageClose zoomPrevious imageIn the NAVSIM test, which simulates complex road situations, QualiDrive achieved 94.5 PDMS in version v1 and represents the latest technology - 92.74 EPDMS in version v2. These results are important not only from a numerical point of view but also from a qualitative perspective. Research has shown that the model not only improved the average trajectory quality but significantly increased the proportion of very high-quality solutions. Compared to an approach based solely on monitoring the final quality, the proportion of trajectories with EPDMS above 0.95 increased from 12.94% to 21.53%, and above 0.98 - from 9.71% to 15.93%.
It is important that the improvement in EPDMS was not dramatic - from 98.52 to 98.57 - but this is due to the fact that the final supervision was already very effective. The difference is that the new approach not only improves the average but also expands the range of good solutions. This means that the system has a greater chance of generating a very high-quality trajectory even in difficult situations - which is crucial for safety and user trust.
What does this mean for the future of autonomous driving?
A new approach to diffusion along the quality axis may have profound implications for the development of autonomous systems. If a model not only generates a trajectory but also develops it step by step from low to high quality, this means greater control over the process and a better understanding of what exactly makes one trajectory better than another. This can lead to more predictable and safer decisions in difficult situations - such as overtaking maneuvers or crossing intersections.
At the same time, it is worth emphasizing that the results are limited to the NAVSIM test and do not yet confirm effectiveness in real road conditions. The model has only been tested on simulations, and its performance in the real world remains unknown. There is a lack of data on computational performance, reaction time or resistance to sensor noise. Therefore, although the results are promising, it is not yet possible to talk about full implementation - this is a step towards a new approach, not a ready-made solution.
Diffusion as a quality process, and not just denoising.
The research results suggest that diffusion in autonomous models is not only a noise removal process - it is a process of gradually improving quality. When the model learns to reverse the degradation of trajectories in an orderly manner, it begins to understand what "good" maneuver means. This changes the paradigm: instead of generating something "less noisy", the system learns to generate something "better". This approach may be key to building trust - if the user knows that the system not only avoids errors but also strives for the highest quality decisions.
However, this does not mean that everything is already solved. The model still relies on data from simulations and has not been verified in real conditions. There is also a lack of information about its scalability, performance or ability to integrate with existing control systems. Therefore, although the new method is promising, its practical application requires further research and testing in the real world - but it already shows that the future of autonomous driving may lie not in greater computing power, but in a deeper understanding of the quality of decisions.