Reactive Planning With Legged Robots In Unknown Environments

Vasileios Vasilopoulos, University of Pennsylvania


Unlike the problem of safe task and motion planning in a completely known environment, the setting where the obstacles in a robot's workspace are not initially known and are incrementally revealed online has so far received little theoretical interest, with existing algorithms usually demanding constant deliberative replanning in the presence of unanticipated conditions. Moreover, even though recent advances show that legged platforms are becoming better at traversing rough terrains and environments, legged robots are still mostly used as locomotion research platforms, with applications restricted to domains where interaction with the environment is usually not needed and actively avoided.

In order to accomplish challenging tasks with such highly dynamic robots in unexplored environments, this research suggests with formal arguments and empirical demonstration the effectiveness of a hierarchical control structure, that we believe is the first provably correct deliberative/reactive planner to engage an unmodified general purpose mobile manipulator in physical rearrangements of its environment. To this end, we develop the mobile manipulation maneuvers to accomplish each task at hand, successfully anchor the useful kinematic unicycle template to control our legged platforms, and integrate perceptual feedback with low-level control to coordinate each robot's movement.

At the same time, this research builds toward a useful abstraction for task planning in unknown environments, and provides an avenue for incorporating partial prior knowledge within a deterministic framework well suited to existing vector field planning methods, by exploiting recent developments in semantic SLAM and object pose and triangular mesh extraction using convolutional neural net architectures. Under specific sufficient conditions, formal results guarantee collision avoidance and convergence to designated (fixed or slowly moving) targets, for both a single robot and a robot gripping and manipulating objects, in previously unexplored workspaces cluttered with non-convex obstacles. We encourage the application of our methods by providing accompanying software with open-source implementations of our algorithms.