Emerging Role of Ultracompact Drones
Ultracompact drones are poised to play a pivotal role in various critical scenarios, including disaster rescue operations, facility inspections, and environmental studies. However, their diminutive size, typically weighing less than 100 grams, imposes stringent limitations on the onboard sensors they can accommodate, their computational capacity, and energy autonomy. Furthermore, conventional sensors such as cameras and LiDAR often fail in challenging environments, such as smoke-filled buildings, dark caves, or locations with transparent or reflective obstacles.
In light of these challenges, a research team from Delft University of Technology (TU Delft) in the Netherlands has drawn inspiration from the tactile whiskers of rodents like rats and moles. The objective is to develop a lightweight, low-latency tactile sensor that can be integrated into drones, allowing them to navigate their surroundings by lightly brushing against objects, thus reducing reliance on visual input.
“Our aim is to equip drones with an advanced tactile sensing capability. This is not about manipulating objects in mid-air, but rather a novel concept known as ‘tactile navigation’, enabling drones to explore unfamiliar environments through physical contact with surrounding objects,” explains Salua Hamaza, an associate professor at TU Delft. “For tactile sensing to be effective in a drone, it must be lightweight, have low latency and consume minimal energy. We found our solution in the natural whiskers of the animal kingdom.”
Innovative Tactile Sensors Mimicking Nature
The research team led by Hamaza has successfully affixed two tactile sensors to the front of a compact drone, roughly the size of a human palm. These sensors are crafted from “nitinol” wire, a nickel-titanium alloy, measuring 0.4 mm in diameter and 200 mm in length. At their base are three small atmospheric pressure sensors; the mechanism relies on estimating the depth and position of contact by analysing pressure changes when the whiskers are flexed. All necessary calculations for detection, decision-making, and autonomous flight control are processed by a microcontroller integrated within the drone, eliminating the need for external computing.
According to the researchers, the angle at which the whiskers are mounted is of particular significance. When navigating alongside a wall, the force exerted by the whiskers can induce rotation in the drone, potentially destabilising its flight. After conducting 72 flights at six different angles across three distinct speeds, they found that the drone maintained stability with whiskers set at angles of 30, 45, and 60 degrees. Ultimately, they opted for a 45-degree angle, as it resulted in the least deviation from the drone’s intended orientation while also noting that the detection range in front of the drone is limited to less than 60 degrees.
A major challenge encountered was the differentiation between genuine contact signals and disturbances caused by the drone itself. The airflow generated by the drone’s propellers can create noise, drift, and other distortions, complicating sensor readings.
To address this issue, the research team developed a correction algorithm that continuously estimates the trend of deviations over time based on sensor values when no contact occurs, effectively filtering out noise caused by vibrations during flight. Consequently, the false positive rate for contact detection when the sensor is untouched decreased from 38.24% (using basic filtering) to 0%.
Enhancing Distance Estimation and Navigation
As part of the process for estimating the distance to nearby walls, the corrected signal is integrated into a neural network prediction through a Kalman filter, a computational method that stabilises estimated values by combining multiple uncertain data sources. This led to a significant improvement in the mean absolute error of one whisker on whiteboard walls, reducing it from 6.40 mm to 4.23 mm, and from 11.33 mm to 5.91 mm on transparent glass surfaces. The tactile sensors have proven to deliver stable distance estimations, unlike laser distance sensors, which can yield inconsistent results when measuring distances to glass walls.
In experimental trials, the drone was tested in an environment containing three transparent walls aligned side by side. When the tactile sensor contacted a wall, the drone maintained a distance of 60 to 100 mm, successfully reaching the target point in all five attempts. It also achieved this consistently when the walls were positioned at angles. In contrast, a laser sensor installed for comparison purposes failed to measure distances unless the drone was directly aligned with the wall.
However, simply navigating along walls can lead to difficulties in locating corners or detecting lateral obstacles. To mitigate this, the researchers devised a search method prioritising the exploration of new areas by probabilistically estimating wall shapes using an implicit surface function based on Gaussian processes. This approach enabled them to locate an exit measuring 80 cm wide in a 2 m square room with glass walls in 9 out of 10 simulations and real-world tests.
Future Prospects and Ongoing Challenges
The microcontroller of the drone has a memory capacity of 192 KB, yet the tactile sensor algorithm employed in this experiment requires only 34 KB. “We aimed to demonstrate that it is possible to achieve tactile feedback without sacrificing the size or processing power of the drone,” asserts Chaoshan Ye, a PhD student at TU Delft. “Even a small drone can detect and respond to its environment in real-time.”
Nevertheless, several challenges remain. Prolonged tracking of a wall can result in cumulative signal deviations, diminishing the accuracy of distance estimations. Additionally, the individual differences in barometric pressure sensors and slight mechanical variations complicate the direct application of a trained model to different sensors. Presently, to achieve optimal performance, an external measurement system is necessary for recalibrating each sensor, which involves tracking the wall 5 to 6 times, taking between 15 and 20 minutes.
The research team envisions applications in humanitarian aid during disasters, exploration in hazardous environments, autonomous flight in confined spaces where conventional sensors may not function, and inspection of complex structures. This research illustrates that by merging lightweight hardware with advanced onboard processing, tactile sensing can provide a robust detection method even for the smallest autonomous flying robots.