Feel the Noise

Robots working everywhere from warehouses and operating rooms to homes have to be able to feel their way around to do much of anything that is really useful. Just imagine how much difficulty you would have in your daily life without the sense of touch. However, giving robots a human-like sense of touch is deceptively challenging. As such, there is no one-size-fits-all solution that works well for every task.

One of the most accurate touch sensing systems available today is GelSight. These sensors use cameras to monitor the interior structure of a flexible material. As the material is deformed through contact with other objects, precise tactile information can be inferred. This approach works exceptionally well for most use cases, but when it comes to agriculture, it often falls flat. All of the branches and leaves occlude the sensor, limiting its accuracy.

Traditional pressure sensors may work better in these cases, but for sufficient resolution, the entire surface area of a robot may need to be covered with them, which would be prohibitively expensive. For this reason, a team at Carnegie Mellon University developed another way for robots to feel their way around cluttered environments. Their system, called SonicBoom, is not fooled by a tangle of branches and leaves, and it does not not require a dense array of expensive sensors.

SonicBoom uses contact microphones embedded in a robot arm to detect sound waves that travel through the structure when it touches something. When the arm makes contact with a branch or other object, the sound propagates through the arm until it reaches the microphones. By analyzing differences in the timing and characteristics of the sound arriving at each microphone, the system can determine the location of the contact with centimeter-level precision.

Unlike camera-based sensors that sit on the exterior of the robot and can be easily damaged or blocked, SonicBoom’s microphones are embedded deeper in the structure, making them more robust and better suited for tough, outdoor conditions. And because only a handful of microphones are needed, rather than dozens or hundreds of pressure sensors, the system is also more affordable and easier to scale.

To train SonicBoom, the researchers collected over 18,000 samples of the robot being tapped with a wooden rod, creating a comprehensive dataset that links specific sound patterns to exact points of contact. The resulting system can detect contact points with a remarkably low error rate of 0.43 centimeters, and even performs well on unfamiliar materials like plastic or aluminum, with just a 2.22 cm average error.

The team is now working on extending SonicBoom’s capabilities so the robot can identify the type of object it touches, whether it is a leaf, a branch, a tree trunk, or otherwise. This could help agricultural robots make smarter decisions in the field, like whether to push through a plant or reroute to avoid damaging it.

While SonicBoom has yet to be tested in real-world farming scenarios, its success in lab settings suggests that it has a promising future. As robots become more heavily used for challenging tasks like agriculture, SonicBoom may help them navigate the world by feel, just like we do.