Perceiving guaranteed collision-free robot trajectories in unknown and unpredictable environments.

摘要

The dissertation introduces novel approaches for solving a fundamental problem: detecting a collision-free robot trajectory based on sensing in real-world environments that are mostly unknown and unpredictable, i.e., obstacle geometries and their motions are unknown. Such a collision-free trajectory must provide a guarantee of safe robot motion by accounting for robot motion uncertainty and obstacle motion uncertainty. Further, as simultaneous planning and execution of robot motion is required to navigate in such environments, the collision-free trajectory must be detected in real-time.;Two novel concepts: (a) dynamic envelopes and (b) atomic obstacles, are introduced to perceive if a robot at a configuration q, at a future time t, i.e., at a point χ = (q, t) in the robot's configuration-time space (CT space), will be collision-free or not, based on sensor data generated at each sensing moment τ, in real-time. A dynamic envelope detects a collision-free region in the CT space in spite of unknown motions of obstacles. Atomic obstacles are used to represent perceived unknown obstacles in the environment at each sensing moment. The robot motion uncertainty is modeled by considering that a robot actually moves in a certain tunnel of a desired trajectory in its CT space. An approach based on dynamic envelopes is presented for detecting if a continuous tunnel of trajectories are guaranteed collision-free in an unpredictable environment, where obstacle motions are unknown. An efficient collision-checker is also developed that can perform fast real-time collision detection between a dynamic envelope and a large number of atomic obstacles in an unknown environment. The effectiveness of these methods is tested for different robots using both simulations and real-world experiments.