ExecDiffuser: Bridging the Gap in Autonomous Driving Execution

Paper: ExecDiffuser: Closing the Sample-to-Execution Gap in Diffusion Planning

Authors: Shuo Pei, Bosun Liang, Saihui Zha, Chen Sun, Huachun Tan, Yong Wang

Executive Summary

Recent research has focused on the challenges of generative planners in autonomous driving, specifically the gap between generating potential driving trajectories and executing a chosen one in real-world scenarios. This study identifies what is termed the "sample-to-execution gap," where simply improving the quality of generated proposals does not necessarily lead to better executed decisions. In fact, it can sometimes make execution worse.

To address this issue, the researchers introduce a new framework called ExecDiffuser. This system integrates three key processes: generating driving proposals, selecting the best trajectory, and replanning based on execution performance. ExecDiffuser employs a unique method known as Intent-Guided Diffusion Proposal (IGD), which assigns responsibility for each hypothesis and shapes the driving paths based on real-time execution feedback. Additionally, it uses Execution-Grounded Selection (EGS) to evaluate the utility of each candidate trajectory, ultimately committing to one path for control while informing future planning cycles with data from the execution.

The effectiveness of ExecDiffuser was tested using three different benchmark environments: nuPlan, InterPlan, and NAVSIM. The framework demonstrated superior closed-loop performance compared to existing leading models, including Diffusion Planner and FlowDrive. Specifically, ExecDiffuser achieved an average score of 85.62 under a unified evaluation protocol, marking an improvement of 4.02 points over Diffusion Planner and 1.31 points over FlowDrive. On the NAVSIM platform, ExecDiffuser secured a score of 92.6 PDMS following the official evaluation pipeline.

The research further validated ExecDiffuser's capabilities through matched counterfactual analysis, which indicated that it reliably transforms high-quality candidate trajectories into effective decisions. This advantage remains evident even after the physical execution of the driving task and subsequent replanning efforts.

The implications of this research are significant for the development of autonomous driving technologies. By bridging the sample-to-execution gap, ExecDiffuser may enhance the reliability of decision-making in autonomous systems. Companies in the automotive and transportation sectors could benefit from this framework, potentially leading to safer and more efficient self-driving vehicles. As the technology matures and is applied in real-world settings, continuous improvements in trajectory planning and execution could be realized, paving the way for more advanced autonomous driving solutions.

Academic Abstract

Generative planners are often evaluated by the trajectories they can generate, although closed-loop driving ultimately depends on the single candidate committed to control. We identify this mismatch as the sample-to-execution gap and show that improving proposal quality alone can worsen the executed decision. Here, we introduce ExecDiffuser, a recurrent proposal-to-execution framework that couples candidate generation, trajectory selection, and subsequent replanning through shared execution semantics. Intent-Guided Diffusion Proposal (IGD) assigns hypothesis responsibility and shapes trajectories with execution-conditioned supervision. Execution-Grounded Selection (EGS) learns a candidate-conditioned publication utility, commits one trajectory to control, and returns it as context for the next planning cycle. Across nuPlan, InterPlan, and NAVSIM, ExecDiffuser achieves leading closed-loop performance against strong state-of-the-art baselines. Under a unified five-seed protocol, it obtains an average score of 85.62, improving over Diffusion Planner and FlowDrive by 4.02 and 1.31 points, respectively. On NAVSIM, it achieves 92.6 PDMS under the official evaluation pipeline. Matched counterfactual analysis further shows that ExecDiffuser more reliably converts strong candidates into committed decisions, while prefix interventions confirm that this advantage persists after physical execution and subsequent replanning.