On 12 August at 12:15, Basant Sharma will defend his doctoral thesis in the field of Computer Engineering “Sample-Efficient Risk-Aware Trajectory Optimization Leveraging Maximum Mean Discrepancy”.
Supervisor:
Professor Arun Kumar Singh, University of Tartu
Opponent:
Associate Professor Johannes Betz, Vice-Rector for Faculty and Research, Technical University of Munich (Germany)
Summary:
This dissertation addresses risk-aware motion planning for autonomous robots operating under uncertainty. The two central challenges are computationally tractable collision-risk estimation (C1) and sample-efficient risk evaluation (C2), particularly when uncertainty is non-Gaussian, multimodal, or available only through samples. The key idea is to formulate collision risk as a distribution-comparison problem. A collision-constraint residual is defined and its distribution is compared against a Dirac delta distribution representing perfect safety using Reproducing Kernel Hilbert Space (RKHS) embeddings and the Maximum Mean Discrepancy (MMD) metric. This enables direct risk estimation from samples without restrictive distributional assumptions.
The first application considers autonomous driving with uncertain obstacle trajectories. The proposed MMD-OPT framework uses trajectory samples and a reduced-set mechanism to achieve lower collision rates than SAA- and CVaR-based approaches in low-sample regimes. The second application extends the framework to trajectory optimization and model predictive control under stochastic dynamics, demonstrating improved safety and fewer constraint violations. The third application focuses on monocular vision-based navigation, where a probabilistic clearance model is learned and integrated with risk-aware control.
The main contributions are: (1) a general MMD-based risk surrogate for nonparametric collision-risk estimation, (2) a reduced-set mechanism for improved sample efficiency, (3) validation across three uncertainty sources—obstacle prediction, stochastic dynamics, and monocular perception—and (4) consistently improved safety in low-sample settings. Overall, the dissertation demonstrates that an RKHS-, MMD-, and reduced-set-based framework can combine theoretical rigor, computational efficiency, and practical applicability for safe decision-making in real-world robotic systems.
Defence can be followed in Zoom: Doctoral Defence (meeting ID: 953 058 8152, passcode: kaitsmine).
