Maravall and de Lope describe a robotic model capable of navigating along aerial power, telephone, or railroad lines, as well as in reticulated structures.
The model begins by schematizing sequential kinematic actions, by humans or primates, as two basic movements: grasp and release. Although bio-inspired, the method does not need to know the empirical positions, or shapes, of obstacles. Instead, to obtain collision-free trajectories, complex senso-motor tasks are simulated by means of a reinforcement learning process, controlled by perceptual feedback. The physical prototype, a robotic mechanism equipped with motorized extremities, namely, three revolute joints, is able to move forward and backward along a horizontal axis. Control of the robot is achieved by computing link torque equations, which generate trajectories for obstacle avoidance. Robot perception, in the sense of physical or engineering intelligence, relies on live images subject to real-time restrictions. The goal of robot survival eschews aprioristic formalization, in favor of if-then-else, or reactive modeling: collision-free trajectories using reinforcement learning paradigms. In brief, computations can replicate functional and physiological animal behavior.
After reviewing issues of configurative space, the so-called global/local dichotomy, the authors opt for a gradient algorithm based on a single-layer perceptron. The migrant robot thus evades obstacles by relying on a computational symbiosis of reinforcement control and perceptual feedback. The main strength of the authors’ approach is that it can be extended to other articulated mechanisms, including manipulator arms.
In Appendix 1, the authors describe a Sharp infrared sensor and a Motorola microcontroller with servomotor. Appendix 2 contains a graph of trajectory control.