Acoustic Communication and Sensing for
Inflatable Modular Soft Robots

The Linbot soft modular platform [13] is the most directly related to this work. It uses a voice coil for actuation, sensing, and communication, taking advantage of the wide frequency response in a similar manner to how we use our piezoelectric transducers. A Hall-effect sensor is used for proprioception through sensing of the voice coil position, electromagnetic coupling between neighboring Linbots allows for omnidirectional communication, and audible range waves can be produced for external communication. To accomplish this they rely on the rigid connections between the actuator core and the extents of the soft shell, only operating with shape changes of up to approximately 30$\%$ on their principle axis. In contrast, our robots undergo maximum volume changes of close to 1000$\%$, and the exterior surfaces do not remain in contact with the primary actuator.

Swarm platforms are relevant in this context because they are also motivated by finding low complexity and cost, scalable solutions [14]. The Kilobot [12] platform uses an IR transmitter and receiver on its underside to both communicate with and detect the distance of neighbors, using only one pair for both functions but doing so only omnidirectionally and only up to about 10cm away. The Open E-Puck platform [15] uses a set of 12 pairs of radially arranged IR transmitters and receivers to perform inter-robot communication as well as range and bearing measurements, which allows it to send and receive signals from specific directions. Our acoustic solution adds the additional functionality of long-range ($>$1m) communication with no line-of-sight requirements, as well as contact/deformation sensing, while only requiring a single transducer instead of an emitter/receiver pair.

A common use of acoustic waves in multi-robot systems is for ultrasonic range estimation. The relatively slow speed of sound lessens signal processing constraints relative to radio frequency solutions (e.g., RSSI) by enabling direct time of flight measurements, making it a useful supplement to improve robustness of distance estimation [16]. Relative positioning of multi-robot systems using ultrasonic ranging at distances up to seven meters has been demonstrated with absolute average error of only 8mm [17].

As opposed to sensing, acoustic communication between autonomous robots is a relatively underexplored area. An exception is in the realm of autonomous underwater vehicles, which are driven towards acoustic modes by the high electromagnetic absorption of seawater [18, 19]. Audible range communication has been noted as a potentially useful supplement to radio frequency networking for land-based multi-robot systems due to the fact that the relatively strong environmental attenuation of acoustic waves can encode environmental information [20]. In this work, we take this idea further by using the soft pressurized structure of the robot itself as the information-encoding transmission environment.

Outside of the robotics domain, acoustic communication has been shown between pressurized mylar balloons that act as amplifiers and speakers when actuated by piezoelectric transducers [21], which served as an inspiration for the communication-at-a-distance in this work.

The inflatable robot units (Fig. 2A) are based on our co-authors’ recent prior work [22] demonstrating untethered cellular robots. Each is composed of a 45cm maximum diameter, 0.4mm thick latex membrane enclosing a DC air pump, solenoid-controlled valve, and up to N (N = 8 given the analog switch array used in this work) acoustic modules distributed across the membrane. Each acoustic module represents a sensing, communication, and connection point for the robots to interact with their environment and each other. There is an inherent tradeoff between the increased functionality (e.g., in terms of sensing resolution) and the increased complexity for each additional acoustic module which bears future investigation.

The acoustic modules (Fig. 2C) comprise FDM 3D printed, cylindrical enclosures (30mm diameter, 7mm thickness, PLA) with 60 degree radially arrayed slots for diametrically polarized cylindrical magnets (3.2mm diameter, 6.4mm height). The magnet housings are slightly over-sized, allowing the magnets to reorient when connectors are drawn together, making them “genderless.” We rely upon passive reorientation of the units for alignment, although notably vibrations can be transmitted with sufficient signal-to-noise ratio through even imperfectly aligned modules. The piezoelectric transducers (27mm diameter brass plate with 20mm diameter ceramic piezo, 0.5mm thickness) are standard contact microphones, fixed into the printed enclosures with 3M 300LSE double sided adhesive tape. The acoustic modules are each fixed to the inside of the latex skin with the same tape.

A Teensy4, which contains a 600MHz Cortex M7 microprocessor, is sufficient for the software-defined radio architecture of the acoustic communication. To minimize cost and complexity the piezoelectric transducers are connected through an 8:16 analog switch array (MT8816) to a single full H-bridge dual-channel motor driver (TB6612FNG) and a single audio amplifier (MAX9814, includes preamplifier, variable gain stage, and output amplifier) with 60dB gain. A block diagram of the system is shown in Fig. 3. Each piezoelectric transducer consumes approximately 60mW during full-duty cycle operation between 3kHz and 20kHz (10mA at 6V). The poor impedance matching between the piezoelectric transducer and the amplifier, which is designed for standard electret condenser microphones, creates a high pass filter around 2kHz. Although here we show units with electronics and power located externally to the robot membrane, prior work shows that the required electronics, battery pack, and charging circuit can be incorporated into the latex membranes inside a 3D printed enclosure [22].

Swarm and modular robot systems designed for large agent counts typically rely heavily on local communication as a way to overcome challenges with scaling of radio-based networks [14]. For modular robots the connection points represent natural avenues for information transfer, such as through direct electrical connections [23]. Methods that do not rely on mechanically flush or material-specific connections, like IR transmit/receive pairs built into the faces of the connectors [24], are more suitable for deformable surfaces. In our robots, the piezoelectric transducers in the rigid enclosure of the magnetic connectors can transfer information in the form of shared vibration through even imperfect contact made between connectors; the received signal amplitude for a 18kHz tone decreases from its full value when all three magnets are aligned, $N_{contacts}=3$, by about 35$\%$ for $N$=2 and 45$\%$ for $N$=1, never falling below about 40dB signal-to-noise ratio (SNR). An advantage over an IR-based method is that vibrations are coupled from the interior modules through the exterior surfaces of the robots mechanically, removing any optical property design constraints.

We implement acoustic communication through binary frequency-shift keying (FSK), chosen over amplitude modulation in order to resist contact-quality based errors. A demonstration of the achievable packet delivery ratio (PDR) for a 1:1 module pair is shown in Fig. 4. Packets consist of a 4-bit start sequence, 4-bit data structure, and one parity bit. The decrease in achievable PDR is correlated with increasingly tight timing requirements (i.e., a shorter symbol time requires stricter phase alignment) and a decreased SNR caused presumably by the piezoelectric transducers being unable to ring up to full vibration amplitude before a bit transition.

Having multiple individually addressable communication points on each robot allows for directional communication between any number of connected neighbors. For this to be possible, signals received at each transducer must be able to be successfully disambiguated from those received at their neighboring nodes; as the signals here are mechanically coupled to the structure and not based on line-of-sight, they radiate symmetrically from their coupling point through the elastic membrane and are received at neighboring points. Fig. 5 shows the received signal amplitude at the receiver node versus the received signal amplitude at the neighboring nodes on both the send and receive robots. Vibrations are increasingly attenuated at higher signal frequencies resulting in a higher SNR in the ultrasonic range. This means that ultrasonic signals are the best option for sending information directionally through the connections, and can do so with the added benefits of being inaudible and having minimal chance of encountering relevant environmental noise.

For multi-connection data routing over a single channel – in this case, an individual robot’s software-defined FSK receiver, which is only hooked up to a single acoustic module at a time – we implement a slotless architecture based on the ALOHA protocol [25]. The default listening behavior is to time multiplex through the $N_{module}$ acoustic modules with an interval equal to a single packet duration $t_{packet}$, waiting to detect a start sequence (0111) and “locking” (i.e., remaining listening) if one is detected. If a full packet is decoded with the correct parity bit, an acknowledgement is then sent through the appropriate module. The corresponding sending behavior is to continuously broadcast a packet on all desired output modules for a duration equal to $N_{module}\cdot t_{packet}$, then listen on those modules for the acknowledgement; if no acknowledgement is received the packet is resent.

Collaboration between our robots is possible without either direct contact or line-of-sight via transmission of signals in the audible range, produced effectively by the same piezoelectric transducers thanks to their flat frequency response. In this case, the pressurized elastic skin acts as an omnidirectional pickup for the airborne acoustic waves, letting the ostensibly contact-based piezoelectric transducers act as true microphones. By operating at the approximate resonance of the piezoelectric transducers of 6kHz signals from robots up to a meter away can be received through the air with a measured SNR of $\approx$7dB through the entire operational volume range ($\approx 0.05-0.5m^{3}$). The received signal amplitude is determined by factors including the robot distance, each robots’ volume, and the contact quality between the acoustic modules and the elastic membrane.

One important and fundamental function of decentralized multi-robot systems is the ability to synchronize in time [26]. In nature, animals use both acoustic and optical (e.g., in katydids [27] and fireflies [28], respectively) signals to achieve synchronicity in a process known as “synchronized chorusing,” or more formally as groups of pulse coupled oscillators.

Here, pulse coupled synchronization using audible signals is implemented simply; an example spectrogram from synchronization of two robots is shown in Fig. 6. Each robot starts with some initial phase offset from its neighbors (about 250ms in Fig. 6). After a delay $t_{a}$, a synchronization pulse is produced by all $N_{module}$ transducers simultaneously at 6kHz for a duration $t_{chirp}$. The cycle repeats after another delay $t_{b}$. During each delay interval a module acting as the receiver is continuously sampled in order to detect amplitude peaks at 6kHz above a predetermined ambient noise threshold. At the conclusion of the $t_{a}+t_{chirp}+t_{b}$ duration cycle, the tallied detections are used to determine whether the chirp should be shifted “forward” or “backward” in a binary fashion; if more are detected during $t_{a}$, for example, then the majority of neighboring robots are pulsing before this one, so the phase is shifted without changing the period by setting $t_{a}\mathrel{{-}{=}}t_{shift}$ and $t_{b}\mathrel{{+}{=}}t_{shift}$.

There is a tradeoff between synchronization time and total (audible) robot count. In the most extreme case, all time slots in the listening period would be filled with chirps and therefore balanced. This means that the time for synchronization is expected to scale with the number of robots as the listening period must increase for additional robots. Time-varying chirps (such as those produced by katydids [29]) could provide an additional layer of information that improves the scalability of this approach.

The synchronization accuracy is related to both the chirp duration and the digital signal processing on the receiver. The minimum chirp duration is bounded by the response time of the piezoelectric transducers and the associated SNR at the receiver side. For the receiver processing, non-overlapping 256-point FFT segments with a sampling frequency of 50kS/s results in a minimum synchronization window of approximately 5ms.

Sensing external stimuli like applied loads and environmental contacts is a critical robotic function. Existing solutions for soft robots, such as adding flexible signal transmission channels (e.g., optical channels [30] or printed traces [31]) throughout the robot surface, are costly, complex, and not robust to the high-percentage shape change exhibited by our robots. In order to sense contact we instead take advantage of the coupling between loads on the robot and the resultant attenuation of the acoustic waves being transmitted through the existing unmodified external surface, reducing instrumentation cost and complexity by taking advantage of the compliant nature of the robot.

Fig. 7 demonstrates that acoustic signals, received at a central receiving node from tones transmitted by surrounding nodes, can be used to detect compression of the robot. Regions are effectively “sensitized” by adding a continuously sampling receiver. Contacts with areas $\geq a_{mod}$ centered on the transmitting modules both dampen the vibrations of the piezoelectric transducer in its magnetic enclosure as well as decrease the coupling of the surrounding elastic membrane to the node, producing a clearly distinguishable shift in received signal FFT amplitude at the tone frequency. The sensitive region size is determined by the initial SNR of the received tones, which is a function of inflated volume, pressure, and contact quality. The spatial resolution is determined geometrically by the acoustic module area, $a_{mod}$, the module dispersion density, and the current inflated volume. In this inverse to the problem of private contact-based communication, it is important to maximize signal transmission to neighboring nodes and hence requires audible-range signals (see Fig. 5). Importantly, contact at the receiver node itself manifests as decreases in amplitude from all surrounded nodes; switching the set of “sensitized” nodes by reconfiguring the analog crosspoint array could allow for diambiguation.