Reference Asset for Full Text 文章配套在线附表

A Physics-Grounded Hierarchical Boolean Decision Framework for Autonomous Navigation 从黑盒到透明：可审计的机器人导航运算框架 TMF

Table 1: Qualitative Comparison of Structural and Safety Attributes Across Four Mainstream Autonomous Navigation Decision Paradigms 表 1：四大自动驾驶导航决策范式的结构性与安全特性定性对比

Paradigm Attributes 对比维度	Pure E2E VLA 纯端到端（VLA）	MindVLA 新一代（MindVLA）	Constraint Hierarchy 约束层级方法	TMF (Our Framework) 可通行介质框架
Explicit Physical/Semantic Boundary	No	No	No	YesExplicit Structural Separation
Source of Explainability	Post-hocApproximation	TextualChain-of-Thought	WeightsCost Function	Architecture-GroundedSystemic Separation
Physical and Semantic Coupling	ExtremeMonolithic Embedding	WeakDecoupling	PartialDecoupling	StrongDeterministic Layer Separation
Emergency Rule Degradation Principle	NoneHigh System Crash Risk	NoneRelies on VLA Generalization	NoneTriggers Global Replanning	Precisely ControlledSemantic Layer Only
Cross-Scenario Transferability	WeakData Distribution Dependent	StrongCommonsense Generalization	WeakRequires Rule Rewriting	StrongPhysics Generalization + Semantic Decoupling
Compliance & Audit Friendliness	Extremely DifficultBlack Box	DifficultHallucination Risks Remain	ModerateHard to Trace Rule Conflicts	StrongAuditable Decision Boundaries
Long-Tail Scenario Management	VulnerableProne to Collisions	ExcellentBut Limited by Latency	Deadlock-ProneOver-Conservative	ExcellentDistinct boundaries via proactive risk compression
Compatibility with Modern Learning Systems	—	Strong	Moderate	StrongAs Deterministic Safety Filter Stream
物理与语义边界显式保留	否	否	否	是显式结构解耦
可解释性来源	事后近似	语言推理链	约束权重代价函数	基于架构的可解释性
物理与语义耦合度	极高混合表征 / 单体嵌入	弱解耦	部分解耦	强确定性分层解耦
紧急规则降级原则	无降级易导致整体崩溃风险	无依赖大模型泛化自适应	无易引发全局重新规划	精准受控仅触碰语义层，物理底线不动
跨场景迁移能力	较弱依赖数据分布	强常识泛化	较弱需重写约束规则	强物理常识泛化与语义解耦
合规与审计友好性	极难完全黑盒	较难推理链仍有幻觉	一般多约束冲突时难追溯	强可审计的决策边界
长尾场景处理	脆弱预判不足易碰撞	优秀但受制于时延与幻觉	易死锁保守策略导致无法前行	优秀边界清晰，通过半径突变法主动压缩风险
与现代学习系统兼容性	—	强	中等	强作为确定性安全过滤流

💡 Architectural Note on Architecture-Grounded Explainability:

We define architecture-grounded explainability as a structural property of the Traversable Medium Framework (TMF). Unlike traditional XAI approaches that seek interpretable representations within unified model blocks, TMF preserves explainability through explicit separation between non-overridable physical safety constraints ($P$) and defeasible semantic constraints ($S$). Navigation decisions emerge from deterministic Boolean operations and Binary Mask Multiplication (BMM), making each intervention structurally traceable and auditable at the architectural level.

💡 关于基于架构的可解释性的核心声明：

我们将基于架构的可解释性定义为可通行介质框架（TMF）的核心结构特性。不同于传统可解释人工智能（XAI）在统一模型块内部寻找可读表征的做法，TMF 的可解释性是通过不可跨越的物理安全约束（$P$）与可放宽的语义约束（$S$）之间的显式隔离来保证的。导航决策完全源自决策层面的确定性布尔运算与二值掩码乘法（BMM），使得系统每一次的行为干预在架构层面均具有强力的结构可追溯性与可审计性。