📝 SummarXiv

Abstract

Many nonconvex problems in robotics can be relaxed into convex formulations via semidefinite programming (SDP), which offers the advantage of global optimality. The practical quality of these solutions, however, critically depends on achieving rank-1 matrices, a condition that typically requires additional tightening. In this work, we focus on trace-constrained SDPs, where the decision variables are positive semidefinite (PSD) matrices with fixed trace values. These additional constraints not only capture important structural properties but also facilitate first-order methods for recovering rank-1 solutions. We introduce customized fixed-trace variables and constraints to represent common robotic quantities such as rotations and translations, which can be exactly recovered when the corresponding variables are rank-1. To further improve practical performance, we develop a gradient-based refinement procedure that projects relaxed SDP solutions toward rank-1, low-cost candidates, which can then be certified for global optimality via the dual problem. We demonstrate that many robotics tasks can be expressed within this trace-constrained SDP framework, and showcase its effectiveness through simulations in perspective-n-point (PnP) estimation, hand-eye calibration, and dual-robot system calibration. To support broader use, we also introduce a modular ``virtual robot'' abstraction that simplifies modeling across different problem settings.