PRIX

Learning to Plan from Raw Pixels for End-to-End Autonomous Driving

Maciej Wozniak¹, Lianhang Liu^1,2, Yixi Cai^{1 *}, Patric Jensfelt¹

¹ Robotics Perception and Learning Department, KTH Royal Institute of Technology ² SCANIA, Stockholm, Sweden (^*) corresponding author, yixica@kth.se

Accepted to IEEE Robotics and Automation Letters (RA-L) 2026!

   

News

• Mar. 3rd, 2026: PRIX is accepted to IEEE Robotics and Automation Letters (RA-L)
• Nov. 19th, 2025: Initial manuscript submitted for review

Table of Contents

• Introduction
• Methodology
• Qualitative Results
• NavSim Setup
• Results & Benchmarks
• Contact
• Acknowledgement
• Citation

Introduction

PRIX (Plan from Raw pIXels) is an efficient end-to-end autonomous driving model designed to operate exclusively on camera data. By eliminating the reliance on expensive LiDAR sensors and computationally heavy Bird's-Eye View (BEV) representations, PRIX addresses the scalability limitations of current mass-market vehicles.

The architecture achieves a state-of-the-art balance between performance and speed, reaching 87.8 PDMS on NavSim-v1 while maintaining a real-time inference speed of 57 FPS on consumer-grade hardware.

Model Performance vs. Speed on NavSim-v1.

Methodology

The core of PRIX is the Context-aware Recalibration Transformer (CaRT), which enhances visual features by modeling long-range dependencies across the spatial domain without explicit 3D geometry. These refined features are then utilized by a Conditional Diffusion Planner. This planner treats trajectory prediction as a denoising process, using a vocabulary of trajectory anchors to refine noisy proposals into safe, feasible paths in just 2 steps.

PRIX Architecture Overview.

Qualitative Results

PRIX demonstrates robust planning capabilities in complex urban environments, safely navigating busy intersections and maintaining performance during adverse weather conditions such as rain and snow.

Robustness in Snow and Rain.

Installation

1. Base Environment (NAVSIM)

PRIX is built on the NAVSIM framework. You must first install the NAVSIM devkit and its dependencies refer to Navsim repo for that .

We also provide environment.yml and requirements.txt file with dependencies.

We recommend using conda and after setting up correct conda env

conda env create --file environment.yml

conda activate prix

pip install -r requirements.txt

NavSim Setup

Evaluation is primarily conducted using the NavSim framework:

• NavSim-v1: Benchmarked in a non-reactive simulation where the agent plans a 4-second trajectory from initial sensor data. Performance is aggregated into the PDM Score (PDMS), which penalizes safety failures while rewarding progress and comfort.
• NavSim-v2: Utilizes pseudo-simulation with reactive traffic, measured by the Extended PDM Score (EPDMS).
• Training: Models are trained for 100 epochs with a per-GPU batch size of 64 using the AdamW optimizer.

Results & Benchmarks

NavSim-v1 Performance

Method	Input	Backbone	PDMS ↑	FPS ↑
UniAD	Camera	ResNet-34	83.4	3
Transfuser	C & L	ResNet-34	84.0	60
DiffusionDrive	C & L	ResNet-34	88.1	45
PRIX (ours)	Camera	ResNet-34	87.8	57

nuScenes Trajectory Prediction

Method	Input	L2 (m) Avg ↓	Collision Rate (%) ↓	FPS ↑
VAD	Camera	0.72	0.22	4.5
SparseDrive	Camera	0.61	0.08	9.0
PRIX (ours)	Camera	0.57	0.07	11.2

Models

PRIX 512 and 256 with ResNet-34 are available on HF https://huggingface.co/maciejw94/prix512/tree/main

Contact

For questions regarding the paper or implementation, please contact Maciej Woznia (maciejw@kth.se) or Yixi Cai (yixica@kth.se).

Acknowledgement

This work was supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP). We also acknowledge the use of the NavSim-v1 and nuScenes benchmarks in our evaluation.

Citation

@article{wozniak2026prix,
  title={PRIX: Learning to Plan from Raw Pixels for End-to-End Autonomous Driving},
  author={Wozniak, Maciej and Liu, Lianhang and Cai, Yixi and Jensfelt, Patric},
  journal={IEEE Robotics and Automation Letters},
  year={2026},
  publisher={IEEE}
}