Deferred-Decision Trajectory Optimization
Purnanand Elango, Selahattin Burak Sarsilmaz, Behcet Acikmese
TL;DR
Deferred-Decision Trajectory Optimization (DDTO) addresses planning under unmodeled uncertainties by maintaining a collection of reachable targets and deferring target selection until more information is available. The authors formulate constrained reachability and show its equivalence to cardinality minimization, enabling tractable optimization-based solution methods. They develop three solution approaches—ddto-qcvx (quasiconvex), ddto-micp (mixed-integer convex), and ddto-scp (sequential convex programming)—that produce trunk-and-branch trajectory structures for discrete-time affine and continuous-time nonlinear systems, demonstrated on quadrotor motion planning problems. The work provides a deterministic, information-gavorable framework with practical relevance to robust autonomous navigation and safe landing in uncertain environments, and outlines future extensions to fully closed-loop perception-driven decision-making.
Abstract
We present DDTO--deferred-decision trajectory optimization--a framework for trajectory generation with resilience to unmodeled uncertainties and contingencies. The key idea is to ensure that a collection of candidate targets is reachable for as long as possible while satisfying constraints, which provides time to quantify the uncertainties. We propose optimization-based constrained reachability formulations and construct equivalent cardinality minimization problems, which then inform the design of computationally tractable and efficient solution methods that leverage state-of-the-art convex solvers and sequential convex programming (SCP) algorithms. The goal of establishing the equivalence between constrained reachability and cardinality minimization is to provide theoretically-sound underpinnings for the proposed solution methods. We demonstrate the solution methods on real-world optimal control applications encountered in quadrotor motion planning.