Towards Trustworthy Autonomous Robots: An Explainable AI-Based Decision Framework
New research proposes TRACE, a decision framework for autonomous robots that links machine actions directly to sensor inputs. The system aims to solve auditability hurdles in deep learning models by building traceable causal chains. How do you audit your deployed models?
Deep learning systems in autonomous robotics present a significant auditing obstacle when incidents happen because investigators struggle to reconstruct the operational logic behind specific choices. A new approach called TRACE introduces a structured architecture designed to connect autonomous actions directly to underlying sensor evidence through documented causal chains. This method divides the decision process into four distinct layers that handle semantic perception, belief reasoning, action synthesis, and execution verification while remaining compatible with deep learning perception models. Testing in simulated warehouse navigation environments demonstrates high performance across metrics measuring evidence traceability, temporal continuity, and decision reconstructability. This architectural development provides a technical pathway to meet transparency mandates established by regulatory frameworks such as the EU AI Act for high-risk deployments.