My current work is focused on developing the mathematical, machine-learning, and robotics foundations required to investigate intelligent autonomous systems, with particular interest in embodied AI, robot learning, computer vision, and multimodal perception.
This page is intentionally sparse. Investigations appear here once they have a research question, a method, and something a reader can check, not before.
Question. Does converting calibrated semantic perception uncertainty into navigation cost improve closed-loop robot outcomes under controlled distribution shift?
A preregistered ROS 2 and Gazebo benchmark testing whether converting calibrated semantic uncertainty into navigation cost makes a Nav2 robot safer under controlled distribution shift. Calibration improved; navigation safety did not.
Calibration improved substantially and clean-route efficiency was preserved, but the preregistered collision-reduction and severity-interaction hypotheses did not pass. Better component calibration did not establish safer navigation.
Question. How early can mobile robot navigation failure be predicted from onboard signals, and can a guarded recovery act on that prediction without degrading outcomes?
A preregistered study of whether navigation failure can be predicted early enough to act on. Failures are detectable roughly four seconds ahead, but the learned predictor did not beat a transparent baseline at a fixed false-alert budget, and prediction-triggered recovery did not improve mission completion.
Both confirmatory hypotheses were not supported. Lead time and calibration held; the comparison against the transparent baseline, generalisation to unseen fault families, and the recovery benefit did not.
Question. If a navigation instruction is treated as an uncertain observation rather than a command, can a robot recover from instructions that are wrong, ambiguous, or contradicted by what it sees?
A preregistered study treating a natural-language instruction as a noisy sensor rather than a command. It closed under a deliberately narrowed scope: a frozen upstream threshold made the preregistered calibration mathematically unreachable, and clean-view grounding saturated at zero errors, which removed the negative outcomes the planned models needed.
The preregistered four-class runtime calibration was not achieved and is not claimed. Two findings stand in its place: the pinned OCR stage truncates score support to [0.5, 1.0], so the required low-confidence bin cannot be populated at all; and human review of all 287 scoreable detections found every error confined to one class, leaving three classes with no negatives to fit.
