We congratulate the following PhD students on successfully defending their theses:
Kang, Hyungsoo
Thesis Title: Coordinated Path-Following Strategies for Multiple Vehicles over Directed Graph Networks
Abstract: This dissertation develops coordinated path-following strategies for fleets of unmanned vehicles operating over general time-varying directed communication graphs. It addresses limitations of prior work that focused mainly on bidirectional graphs by introducing methods applicable to broader digraph settings. A key contribution is an integral connectivity condition, which requires connectivity only over time rather than at every instant. Under this weaker condition, the proposed strategy can still achieve coordination objectives with exponential stability. The work also studies a leader–follower framework in which only leaders know the desired mission rate. Followers estimate this rate using an integral-based protocol based on locally available information. To reduce communication demands, an event-triggered strategy is proposed in which agents transmit data only when necessary. Finally, the dissertation presents a distributed optimal defensive trajectory planning approach for defender UAVs, combining Bernstein approximation and Nash equilibrium–seeking with receding horizon updates to counter adversarial swarm attacks.
Gu, Yuliang
Thesis Title: Context-Aware Safe Autonomy For Decision-Making Under Uncertainty
Abstract: Autonomous systems are increasingly deployed in environments where dynamics, task geometry, and safety conditions change over time and are only partially observed. In such settings, failure arises not only from noise or model mismatch, but from the fact that the appropriate decision depends on latent or changing context. This dissertation develops a framework for context-aware safe autonomy that is organized around two claims: uncertainty should be converted into structured runtime objects that can be inferred online, and those objects should be used at the layer of the decision stack where they first become decision-relevant.
The dissertation develops this principle across the model, planner, policy, and safety layers. At the model layer, Proto-MPC introduces an encoder–prototype–decoder residual dynamics architecture that augments nominal physics with compact context-conditioned corrections for efficient online quadrotor adaptation under changing winds. At the planner layer, BC-EvoCEM improves multimodal stochastic trajectory optimization through performance-weighted Bregman centroids and trust-region respawning. At the policy layer, BCPO formulates contextual reinforcement learning as a coupled inference-and-control problem and learns latent codes that are compact yet control-sufficient. When uncertainty is safety-critical, SL1-Simplex provides verified fallback through switching adaptive control and finite-time model learning, while AutoSafe embeds safety into policy parameterization through smooth composition with a safe prior to support online learning under hard constraints.
Taken together, these contributions support a unified Learn–Infer–Act–Safeguard view of reliable autonomy: learn structure offline, infer the relevant runtime object online, act through the layer where that object matters most, and safeguard behavior when uncertainty becomes high-consequence. The resulting framework shows how uncertainty can be represented in forms that are computationally tractable, decision-relevant, and actionable at runtime.
Tao, Ronald
Thesis Title: High-Performance Model Predictive Control of Uncertain Systems with Safety Guarantees
Abstract: Model Predictive Control (MPC) is a powerful framework for controlling constrained dynamical systems, but its real-world deployment is limited by uncertainty in system dynamics and mission requirements. This dissertation develops methods to enhance both performance and safety of MPC by addressing these two layers of uncertainty. For dynamics uncertainty, a robust adaptive MPC framework integrates L1 adaptive control and robust feedback to compensate matched and attenuate unmatched uncertainties. For mission-level uncertainty, a backup plan safety framework is proposed, enabling systems to maintain feasible alternative missions through a multi-objective multi-horizon MPC formulation. Finally, DiffTune-MPC introduces a differentiable approach for automatically tuning MPC cost parameters based on closed-loop performance. Together, these contributions improve robustness, safety, and usability of MPC in uncertain environments.