Publications / TRACE
Towards Trustworthy Autonomous Robots: An Explainable AI-Based Decision Framework
TRACE, a Transparent Reasoning Architecture for Credible Execution.
, “Towards Trustworthy Autonomous Robots: An Explainable AI-Based Decision Framework,” in SoutheastCon 2026, IEEE, Feb. 2026, pp. 1–6, doi: 10.1109/SoutheastCon63549.2026.11476455.
BibTeX
@inproceedings{Temel2026TRACE,
title = {Towards Trustworthy Autonomous Robots: An Explainable AI-Based
Decision Framework},
author = {Temel, Cagri},
booktitle = {SoutheastCon 2026},
pages = {1--6},
year = {2026},
month = feb,
publisher = {IEEE},
doi = {10.1109/SoutheastCon63549.2026.11476455}
}
The problem
A warehouse robot stops unexpectedly and a worker collides with it. Investigators want to know why the robot chose that action. The network processed sensor data through millions of parameters to produce a stop command, but the chain of evidence leading to that choice was never preserved, so the answer can only be reconstructed speculatively. This gap has three practical consequences: incident investigation becomes guesswork, safety certification under standards such as ISO 13482 becomes difficult because opaque systems cannot produce documented risk assessment, and the EU AI Act requirement that high-risk systems keep logs enabling their functioning to be traced is one most deployed robotic systems cannot satisfy.
The approach
Post-hoc explanation is the wrong shape for this problem. LIME and SHAP explain individual predictions but cannot reconstruct the temporal sequence of decisions leading to an incident, and attention weights are known to be an unreliable proxy for causal influence. TRACE therefore does not extract explanations from an opaque model. It embeds the audit requirement into the decision architecture itself, so that every action is already traceable through documented causal chains to specific sensor evidence. The framework stays model-agnostic: CNNs, transformers, and future architectures can sit inside the layers while auditability is preserved at the decision level.
| Layer | Function | Explanation artifact |
|---|---|---|
| 1. Semantic Perception | Sensor fusion into entity recognition, where each detected entity keeps explicit links to the sensor evidence supporting it | Evidence Aggregation Record |
| 2. Belief Reasoning | Scene graph into probabilistic state estimation | Causal graph |
| 3. Action Synthesis | Constraint-aware planning and decision selection, documenting why an action was chosen over the alternatives | Counterfactual Decision Tree |
| 4. Execution Verification | Command issue with compliance monitoring | Audit trace |
The artifacts combine into a Trust Composite Score in [0, 1]. Counterfactual Decision Trees are the contribution that makes “what if” analysis possible during incident review, which is what an investigator actually needs and what feature attribution cannot give.
Results
Evaluated on warehouse robot navigation across 500 simulated decision cycles.
| Metric | What it measures | Result |
|---|---|---|
| Evidence Traceability | Sensor-to-decision linkage | 98.6% |
| Temporal Continuity | Audit trail completeness | 99.0% |
| Decision Reconstructability | Post-hoc analysis capability | 98.1% |
| Computational overhead | Cost of maintaining the audit structure | sub-millisecond |
The three metrics are themselves part of the contribution: they define what it means for a decision to be reconstructable by an independent investigator, which is a property the explainable-AI literature usually leaves informal.
Related work of mine
TRACE covers auditability after the fact. CT-SAFR covers verification before the act, for robots whose plans come from a language model. Together they are the two halves of the same commitment: an autonomous action should be checkable in advance and reconstructable afterwards.