Introduction
While most mobile robots rely heavily on vision, cameras have inherent limitations: they have a restricted field of view, fail in poor lighting, and cannot "see" through or around solid obstacles. Audition (hearing) provides a powerful complement because it is omnidirectional and functions perfectly in total darkness. However, robotic audition faces two major hurdles: reverberation and spatial ambiguity. In typical indoor environments, sound bounces off walls and ceilings, creating echoes that can confuse localization algorithms. Furthermore, using only two microphones leads to the "cone of confusion," where a system cannot easily distinguish front from back or determine elevation.
The Echo-Avoidance (EA) Model
The core contribution of this research is the Echo-Avoidance (EA) model, a bio-mimetic approach inspired by the human precedence effect. This effect allows humans to prioritize the first-arriving sound (the direct path) and ignore subsequent reflections.
The EA model uses a feedback loop to estimate the "pattern of echoes" and suppresses the localization mechanism during periods dominated by reflections, focusing instead on "echo-free onsets." As shown in the block diagram, the system processes localization cues while simultaneously estimating echoes to generate an inhibition signal $I_e(f, t)$. This signal modulates the integration of localization cues, ensuring that only reliable, direct-path information is used.
The echo estimation process, detailed in Figure 2, models how received sounds generate a decaying pattern of estimated echoes. By comparing the current signal $s(f, t)$ with the estimated echo $s_e(f, t)$, the system can identify onsets where the sound-to-echo ratio is high. This mechanism effectively filters out the "muddiness" of reverberant rooms.
Hardware and Localization Strategy
To solve 3D localization, the robot utilizes a four-microphone array: three arranged in a triangle on a horizontal plane and one on top of a spherical "head." This geometry allows the system to resolve azimuth (horizontal) and elevation (vertical) simultaneously.
Instead of simple cross-correlation, the system generates Arrival Time Difference (ATD) histograms for different frequency components. These are mapped into a 2D azimuth-elevation space. By taking the geometric average of these histograms across different microphone pairs, the robot can identify multiple concurrent sound sources as distinct peaks. This approach favors the efficiency of biological strategies, such as those of the barn owl, which use specific cues for specific spatial dimensions.
Experimental Results
Experiments were conducted in both an anechoic chamber and an ordinary office room. In the anechoic chamber, the system achieved high precision, with a maximum absolute error of only 4 degrees for horizontal sources. In the reverberant office environment, the EA model proved robust against real-world echoes and background noise from fans and air conditioners (up to 40 dB).
A key finding was the robot's ability to navigate toward "invisible" targets around corners. Because sound waves diffract around edges, the robot initially localizes the edge of the doorway. This "minimum path" guidance effectively leads the robot to the opening, where it can then re-localize the true source. The system also successfully distinguished two concurrent speakers talking at different locations (e.g., 0° and 38° azimuth).
Conclusion and Future Work
This bio-mimetic approach demonstrates that prioritizing "echo-free onsets" is a powerful strategy for robust robotic audition. While the current model uses learned parameters, future work aims to make the EA model fully adaptive to rapidly changing acoustic environments. Additionally, integrating wideband signals will be necessary for more complex real-world navigation tasks where phase determination is more challenging.
Summary updated