TU Delft Bird Drone Feels Its Way Onto Branches and Rests Quietly

Photo credit: TU Delft
Forest monitoring usually means a small quadcopter hanging in place while its motors chew through a battery in a few dozen minutes. Noise follows the same schedule. A machine that can grab a branch and cut power would last far longer and bother far fewer animals, yet the last few inches of that landing have wrecked plenty of earlier designs.

Cameras work fine until the gripper enters the frame and the drone is left scrambling to grasp the exact branch it need. Leaves and trees, as well as an overgrown mass of branches, obstruct the view. Salua Hamaza is quite forthright about the issue. Vision brings you close, but when it really counts, your own grip gets in the way. Birds have already figured this out, for example, taking that final stretch with their feet, physically feeling their way along the wood as they gain traction. Hamaza believes that a flying robot would benefit greatly from the same closed loop of feedback.

Hamaza, Anton Bredenbeck, and Anish Jadoenathmisier designed a lightweight hand for their quadrotor with three fingers in about human proportions. Each finger is made up of three parts: a 3D printed PLA spine and a soft silicone pad that can rub against a variety of surfaces, including bark, a wooden beam, a T junction of wood, and even an arm. The fingers spring shut on their own, so holding wood requires almost no extra effort once everything is in order. When the drone has to search, a single tendon opens each finger, and then there are the sensors, which are 9 copper electrodes down under each pad that provide a simple yes-or-no response if they make touch. The foil on the tips of the fingers even gives you a bump on both sides. An MPR121 controller converts the capacitance change into a yes/no signal that the flight computer may use right away.

Even so, the drone starts by guessing with its cameras. But after that, it begins flying in a figure eight pattern, opening and shutting its hand at each end of the loop. The initial touch is all it requires. That’s because the machine already knows the geometry of its own fingers, so a single sensor reading corresponds to a point in space and a direction to move in. Bredenbeck explained it a little more simply: each sensor merely indicates that something is present, but if you know the form of the hand, you can determine where the branch is and how it is pitched. The drone continues to rotate and slide until all three fingers come into contact, at which point it turns off the motors completely. If everything goes wrong and it can’t get a grip, it just returns to a safe hover and tries again.

Simulation runs performed much better than expected, clearing more than 99% of landings even when the starting guess was significantly off, by 60 cm and 50 degrees. Despite the same shaky projections, hardware flights spanning 26 testing were also a success. Then they attempted a version without the tactile feedback loop, which you’d expect to be a disaster, but it actually kept getting closer even when the estimate was off by up to 10 cm. The diameter of the branch it could pick up increased from approximately 10 cm to a lot more respectable 15 cm, all because it could use its fingers to feel its way in. However, it is not perfect, as very thin sticks under 2 cm just glide through without making enough contact to be picked up. One of the cases they observed recognized the first bump at 15 seconds in and locked onto the branch about 35 seconds later. There is a catch if your fingertip becomes stuck on a branch toward the end.
[Source]