Interaction-Aware Trajectory Optimization: Perception and Contact in Autonomous Systems

Abstract

The capabilities of autonomous systems have been challenged in recent years by the complexity of environmental interaction combined with agile motion demands. Such tasks require reasoning about high-dimensional parameterizations of the physical world, and are critical to accomplish meaningful interaction-driven autonomy. This dissertation develops modeling and optimization techniques for three fundamental challenges in interaction-aware trajectory optimization: active spatial perception of unknown environments, decision-making under unmodeled environmental uncertainty, and multibody contact interaction. In each case, the interaction requirement is embedded directly in the continuous-time optimal control problem and solved by sequential convex programming, in most cases at real-time or near-real-time rates. First, we address the set-based viewplanning problem, in which a sensor field of view must contain an extended region of interest, and prove conditions under which the resulting infinite-dimensional constraint admits an exact finite-dimensional representation. Next, we lift the overconstrained structure of multi-landmark viewplanning by embedding a differentiable continuous-time filtering model directly into the optimal control problem, yielding an active simultaneous localization and mapping formulation that jointly optimizes vehicle motion and estimation uncertainty. Then, we introduce a graph-based framework which poses multi-segment trajectory optimization over a directed acyclic graph of trajectory segments, and use it to solve general nonconvex deferred-decision trajectory optimization problems (maintaining reachability to multiple candidate targets for as long as possible) more than an order of magnitude faster than the prior state of the art. An adaptive algorithm is introduced to operationalize this capability for perception-driven safe landing. Finally, we extend continuous-time constraint satisfaction to multibody contact by introducing integral cross-complementarity constraints, enabling contact-implicit trajectory optimization that proposes contact sequences without a prescribed mode schedule. The resulting methods are validated on rocket-powered descent, quadrotor landing and legged locomotion tasks, in both high-fidelity simulators and hardware flight tests, along with consistent benchmarking against relevant baselines at each step.