Event
Ph.D. Dissertation Defense: Clinton Enwerem
Tuesday, November 24, 2026
11:30 a.m.
AVW 1146
Announcement: Ph.D. Dissertation Defense
Name: Clinton Enwerem
Committee:
Professor (Chair): Prof. John S. Baras
Professor: Prof. Calin Belta
Professor: Prof. Dinesh Manocha
Professor: Prof. Michael Wilson Otte
Professor (Dean's Representative): Prof. Nikhil Chopra
Date/time: Tuesday, November 24th, 2026 at 11:30 AM
Location: AVW 1146
Title: Risk-Sensitive Optimization for Robust Autonomy with Applications in Motion Planning & Dexterous Robotic Manipulation
Abstract: Autonomous systems are increasingly deployed in safety-critical settings where even rare failures can have severe consequences, from ground vehicles traversing hazardous unstructured terrain, through autonomous aerial robots inspecting complex infrastructure under degraded sensing and localization, to manipulators performing contact-rich tasks in human environments. Conventional formulations of learning, planning, and control in such settings optimize expected cost. However, because the expected value operator weights adverse outcomes by their probability, rare but severe failures contribute little to the optimization objective. Chance-constrained formulations address this limitation by bounding the probability of constraint violation, but leave the severity of violations unconstrained, offering limited protection against infrequent but catastrophic outcomes. These limitations are particularly consequential in contact-rich manipulation, where small deviations in friction, object pose, or contact geometry can fundamentally alter the forces a grasp can sustain. Consequently, nominal performance measures (e.g., grasp quality) provide little indication of the likelihood or severity of failure under uncertain contact conditions. This thesis develops risk-sensitive methods for autonomy under uncertainty, guided by the principle that decisions must account for adverse outcomes alongside expected cost and nominal performance. We apply this principle to learning, planning, and belief-space control, where uncertainty shapes future costs and constraint violations, and to robotic manipulation, where uncertain contact conditions affect grasp stability and execution.
We first develop a risk-regularized action-value iteration algorithm that incorporates the Conditional Value-at-Risk (CVaR) of a learned violation-cost distribution into a quantile-regression loss. We prove that its distributional Bellman operator is a contraction with a unique fixed point. In reach-avoid navigation, the learned policy succeeds in every test episode, versus half for a risk-neutral baseline. For planning, we bias node expansion in a sampling-based planner toward path segments with low tail cost, reducing failure rates to 2-8% versus 10-16% for the baseline over 50 trials at each noise level. For partial observability, we develop a risk-constrained belief-space controller and establish a lower bound on safety probability under exact satisfaction of its tail-risk constraint. In vision-guided stowing, it achieves 82% success without contact violations, versus 55% and 50% for risk-neutral and chance-constrained baselines that produce nonzero exterior contact forces.
For dexterous manipulation, we address uncertainty in friction and object pose. We construct a risk-adjusted grasp wrench space (GWS) by scaling tangential wrench contributions according to adverse-tail friction. The radius of the largest ball inscribed in the risk-adjusted GWS defines a margin quantifying the wrenches a grasp can resist under uncertainty. We prove that this margin is non-increasing with confidence level and that a positive margin guarantees force closure with at least the prescribed probability for coherent risk measures dominated by Value-at-Risk. Across three dexterous hands, grasps synthesized with our margin achieve 70% lift success under adverse friction, versus 25% using the standard Ferrari-Canny epsilon margin. To complement friction-robust assessment, we develop a structured grasp generator that predicts wrist poses, hand configurations, fingertip contacts, and contact forces from object point clouds. Its SE(3)-equivariant wrist-pose flow handles rigid object transformations, while contact-conditioned force prediction and wrench-balance supervision promote force consistency. Under external disturbances and common geometric refinement, grasp retention reaches 82.1% versus 75.4% for a state-of-the-art baseline. Physical grasp-and-lift experiments further validate the generator.
Realized contacts may differ from those assumed during grasp synthesis, reducing grasp quality. We introduce grasp distance fields (GDFs), smooth functions over coupled arm-hand configurations and a set of grasps satisfying the force-closure criterion. Their negative gradients guide the arm-hand system toward these grasps without separate planning or discrete grasp selection. We enforce collision avoidance, workspace constraints, and clearance using a control-barrier-function (CBF) quadratic program (QP). After grasp establishment, we add a wrench-quality barrier constraint to the QP to limit subsequent margin loss relative to its value at establishment. A dynamics test shows that a grasp rejected by the constraint loses the object under gravity. In bimanual manipulation with dynamic obstacles, our GDF-guided, velocity-aware controller maintains clearance where pose-only feedback and open-loop baselines fail.
Collectively, these results connect risk-sensitive decision-making with probabilistic grasp assessment, pose-consistent grasp generation, and safety-constrained manipulation.
