Overview
This paper introduces an architectural design for autonomous driving Electronic Control Units (ECUs) intended for a transitional period where production vehicles utilize modular multi-neural network (NN) pipelines. The proposed architecture supports the parallel operation of both a modular pipeline and an end-to-end (E2E) path, facilitating a staged migration towards E2E autonomous driving.
Research Context
While end-to-end autonomous driving is identified as the dominant research direction, current production vehicles continue to rely on modular multi-NN pipelines. The presented architecture addresses the challenge of integrating these two approaches within a single system, particularly in the context of production System-on-Chips (SoCs).
The research notes that an "egalitarian late fusion" approach, when transplanted to a production SoC, can be computationally inefficient and lacks a natural unit for the staged substitution of E2E components.
Approach
The proposed architecture is based on three design principles:
- Primary-Path: This principle involves explicitly selecting a primary perception chain. The enclosure of this primary chain within a single SoC pair is prioritized over non-critical paths.
- Exposure-Time-Aligned: This principle propagates the primary sensor's exposure time, denoted as $\tau_{\rm \exp}$, as a tag throughout the processing chain. Fusion nodes are event-driven based on matched $\tau_{\rm \exp}$ values rather than operating on a fixed cycle.
- Co-Path Coexistence: Building upon the first two principles, this allows an E2E output path to run concurrently with the modular pipeline within the same $\tau_{\rm \exp}$ cycle.
The implementation of this architecture is described for a Dual-SoC production Autonomous Driving ECU.
Findings
On a Dual-SoC production AD-ECU, the implementation of the proposed architecture:
- Closed the camera-shutter to planner-output latency at a mean of 296 milliseconds.
- Achieved this latency within a 350-millisecond design budget.
Under the 'Co-Path Coexistence' principle, the modular pipeline functions as the primary system at the production launch. The E2E path operates as a shadow system on real vehicles during this phase. As evaluation evidence accumulates, the scope of the E2E path is designed to be expanded.