Towards reactive whole-body motion planning in cluttered environments by precomputing feasible motion spaces

Abstract

To solve complex whole-body motion planning problems in near real-time, we think it essentials to precompute as much information as possible, including our intended movements and how they affect the geometrical reasoning process. In this paper, we focus on the precomputation of the feasibility of contact transitions in the context of discrete contact planning. Our contribution is twofold: First, we introduce the contact transition and object (CTO) space, a joint space of contact states and geometrical information. Second, we develop an algorithm to precompute the decision boundary between feasible and nonfeasible spaces in the CTO space. This boundary is used for online-planning in classical contact-transition spaces to quickly prune the number of possible future states. By using a classical planning setup of A* together with a l2-norm heuristic, we demonstrate how the prior knowledge about object geometries can achieve near real-time performance in highly-cluttered environments, thereby not only outperforming the state-ofthe-art algorithm, but also having a significantly lower model sparsity.

Multicontact Locomotion of Legged Robots

“…In all these cases, the set boundaries are represented by very simple mathematical structures (typically linear inequalities) in order not to burden the OCP solver. Remarkably, there are few papers about the automatic synthesis of the set boundaries [15], [19], [20].…”

“…The decision variables of problem (20) are the trajectory in state x and control u spaces. The study of the geometrical properties of its solution, through the optimality principles of Hamilton-Jacobi-Belman or Pontryagin Maximum, seems yet out of reach (because of the bilinear constraint (20b)).…”

“…Footstep limits have been defined from data, for example using hyper-planes based on a dataset of robot success and failure inside a dynamic simulator [19]. Similar constraints can be obtained by training a neural network [20]. In [15], bounds of the capturability regions are obtained by extensive computations of the viability set of reduced models.…”

Learning Feasibility Constraints for Multicontact Locomotion of Legged Robots

“…In all these cases, the set boundaries are represented by very simple mathematical structures (typically linear inequalities) in order not to burden the OCP solver. Remarkably, there are few papers about the automatic synthesis of the set boundaries [23,21,32].…”

“…Footstep limits have been encoded by hyper-plane based on a dataset of robot success and failure inside a dynamic simulator [23]. Similar constraints can be obtained by training a training a neural network [21]. In [32], similar bounds are obtained by trial and errors based on stability analysis of the whole-body system.…”