Publications

Core evidence for a preregistered ROS 2/Gazebo/Nav2 study of early navigation-failure prediction and guarded recovery: every report and manifest, the frozen configurations, and all trained checkpoints with their normalisation and training records. Both confirmatory hypotheses were not supported.

The confirmatory core dataset for the risk-calibrated semantic navigation study: 6,840 episode summaries across four navigation systems under controlled distribution shift, with the manifest, exclusion ledger and release inventory needed to reproduce every reported result. Archived separately from the approximately 1 TB retained bag corpus.

A preregistered ROS 2 and Gazebo benchmark testing whether calibrated semantic perception uncertainty, converted into navigation cost, improves closed-loop outcomes under controlled distribution shift. Calibration improved and clean-route efficiency held; the preregistered collision-reduction and severity-interaction hypotheses did not pass.

A preregistered ROS 2/Gazebo/Nav2 study of whether mission-ending navigation failure can be forecast early enough to support a bounded recovery, and whether acting on the forecast improves the mission. Warning was timely where it occurred, with a median useful lead time of 3.8 s, but the primary recall interval included zero and guarded recovery did not improve completion. Both confirmatory hypotheses are unsupported, leaving a signal-to-decision gap.

A leakage-safe reproduction of a widely replicated eye-state and temporal-alerting pipeline under a subject-independent protocol. An oracle-classifier decomposition locates the residual false alerts in the temporal decision rule rather than in classification error, above a threshold of classifier competence.