Sound-based Motion Planning for a Mobile Robot

Any motion-planning problem has two distinct parts: Target specification and the computation of a collision-free path through the workspace from the initial position to the target. From a formal standpoint, the motion planning problem is usually formulated in a geometric framework which assumes that the target position is known. However, in practical applications, this assumption may not be invalid. Most importantly, for most autonomous systems, the goal may be known from multiple inputs: vision, hearing, smell etc., which may only convey its approximate location. Out of these, vision is arguably the most important. However, sound presents a very useful input because of its capability to diffract around obstacles. Consequently, targets which are otherwise invisible may be tracked and located using a combination of audio and visual clues, which would have been impossible in the absence of auditory clues. Its interesting to note that this is the major reason for the evolution of sound perception in mammals since sound is indispensible when a prey/predator is outside the visual range. This project is aimed at investigating the practicability of combining sound and visual navigation for a mobile robot (specifically, the Sony AIBO).

There has been a lot of work on sound localization using an array of audio sensors (microphones) [1,2,3]. These methods aim to find the spatial location of a sound source with respect to the robot (sound localization) using the difference in sound arrival times between pairs of microphones. It is important to note that any reference to spatial localization here means finding the direction of a source only and not its distance. Interestingly, most mammals are able to exercise good sound localization using only two audio sensors as opposed to a whole array. In recent studies described in [4,5], the authors describe a new technique for sound localization which works using only two audio sensors placed antipodally on a spherical "head". Quoting from the paper - "In nature, directional acoustic sensing evolved to rely on diffraction about the head with only two sensors - the ears. The impinging sound waves are modified by the head in a frequency and direction dependent way, and additional complex filtering is performed by the external ears. The cochlea decomposes the sound pressure signal into frequency bands. The brain then uses interaural differences in phase (IPD) and intensity level (ILD) in the various frequency bands to infer the location of a source." It is also important to note that tracking a moving sound source is much more complex than a stationary one. In this project, I will concentrate on handling only fixed targets. Most of the sound localization systems to date (including [4]) assume a relatively obstacle-free environment or a robot with a limited visual capability. In this project I wish to explore scenarios in which both audio and vision play an integral role in motion planning for a mobile robot. Little research has been done in the area of combining auditory inputs with traditional motion planning techniques.

Explore ways to combine the sound localization scheme mentioned in [4] with simple visual motion planning algorithms to make AIBO capable of reaching a target from the initial position under the following scenario:

1. The goal is something which the AIBO can visually "recognize". The goal periodically emits a characteristic sound, which cannot easily be mixed with background noise.

2. The goal is not directly visible to AIBO.

3. AIBO must navigate around obstacles and walk in the general direction where the sound is coming from, even though the goal is not visible, until it catches sight of the goal. This would convincingly demonstrate that it is actually able to find a goal while going around obstacles, based solely on its sound.

 

The basic project goals may be broken down into the following stages:

    AIBO must be taught to "recongnize" sounds. Specifically, it should be able to filter the "goal sound" out of a lot of background noise.

    Use the techniques described in [4] to program the AIBO so that when the goal sound is played repeatedly, it turns to face in the direction of the sound. This would test that sound localization is working.

    Implement basic motion planning algorithms so that the AIBO uses the goal sound to fix the general direction it should be going in, while avoiding obstacles seen using the visual system/range sensors. This would involve a lot of practical issues as the sound localization system will be put in a lot of stress due to noise, especially that emitted by AIBO's motors.

 

Make AIBO capable of tracking and following a sound source moving at a small speed.