TRACE and UDV: Technical Analysis of an Interpretable Perception-Decision Framework for Autonomous Driving

The av-perception-trace project proposes the TRACE structured reasoning protocol and UDV (Understanding-Decision-Verification) reasoning loop, constructing a complete interpretable framework from perception to decision-making. It enables auditable and verifiable autonomous driving decisions, addressing the "black box" dilemma of traditional systems.

Autonomous driving technology is developing rapidly, but the traditional pipeline (sensor → deep learning → control output) forms a "black box". The lack of interpretability in decisions leads to problems such as difficulty in determining accident liability, hard-to-diagnose system anomalies, and no basis for regulatory approval.

The TRACE protocol explicitly converts implicit reasoning into structured symbols, including five elements: Targets, Relations, Action, Constraints, and Explanation. It records clear reasons for decisions (e.g., "Pedestrian on path + Pedestrian priority → STOP").

UDV breaks down decision-making into three stages: Understanding (identifying significant objects, risks, and uncertainties), Decision (actions based on rule constraints and confidence levels), and Verification (consistency checks, counterfactual reasoning, and proof of trajectory), ensuring rigorous decision logic.

The architecture includes a perception layer (extracting object, map, and motion information), TRACE teacher rules (explicit coding based on regulations), a learning factor model (data-driven mapping), a UDV reasoner, and an evaluation report module, forming a complete process.

In a case, the CAN bus showed the vehicle was driving normally, but the perception layer detected constraints such as a pedestrian on the path and an approaching vehicle ahead. TRACE/UDV decided to STOP, reflecting the "safety first" design (prefer false positives over missing dangers).

Advantages: Interpretability (transparent decisions), verifiability (logical inspection), auditability (complete evidence chain), safety (rule constraints); Challenges: Incomplete rule coverage, propagation of perception errors, high computational overhead, and need for optimized human-machine interaction.

This project represents a shift from "black box" to "white box". Its design philosophy emphasizes that autonomous driving needs to "be able to explain decisions", which is core to gaining public trust and regulatory approval, and provides an open platform and toolset for researchers and engineers.