Drones can do a lot of things, but, until now, they have not been able to stick the landing on a ceiling, making upside-down and other off-kilter surfaces off limits for dynamic perching by flying robots. Researchers at Penn State University, however, developed a remarkably versatile biomimetic protocol for teaching drones to flip and perch, based on a deep analysis of how common flies make sticking the (upside-down) landing look easy.
A paper in IEEE Transactions on Robotics (T-RO) by a team led by Bo Cheng, Ph.D., Associate Professor of Mechanical Engineering at Penn State, details how reinforcement learning based on two kinds of visual cues can enable drones to approach, invert, and perch “on the fly,” as it were, in virtually any scenario or surface orientation.
Recently, investigators developed deus ex machina approaches to getting drones onto sloped surfaces — using cables, specific environmental featuresor pre-defined surface inclinations, for example.
Others have employed acrobatic maneuvers that rely on model-based control, with execution in an open loop, which makes the maneuver—and consequently the success of the landing—extremely sensitive to modeling errors and external disturbances.
Cheng and his team took a different, biomimetic approach, combining insect-like landing designs with online feedback control, obviating the need for special prosthetics, limits or highly controlled circumstances. In fact, their work generalizes landing processes to a wide range of inclined surfaces, especially upside down on ceilings.
Marco Tognon, co-chair of the IEEE Technical Committee of Aerial Robotics and UAVs, said the sky’s the limit for drones with such capabilities. “It opens entirely new possibilities for persistent and energy-efficient operation in environments where conventional flight or landing is impractical,” he said. Such autonomous drones “Could perch beneath tree canopies to monitor forest ecosystems, attach to tunnel walls to perform air-quality or traffic measurements, or rest under urban infrastructure such as streetlights and bridges for surveillance and inspection” added Tognon, a researcher with INRIA, the National Institute for Research in Digital Sciences and Technologies in Rennes, France.
Cheng and Penn State doctoral researcher Bryan Habas, the first author of the paper, used high-speed video footage of flies approaching, re-orienting and landing on inverted surfaces in mere milliseconds to develop a two-tiered “universal policy,” an algorithmic cascade that can be generalized to different conditions and orientations, for teaching drones to do likewise.
They discerned three critical aspects of insect perching behaviors that can inform drone engineering:
Using a Gazebo physics environment and a palm-sized Bitcraze Crazyflie 2.1 drone, plus 3D printed legs, the team focused their efforts on the first element — approaches – with a special focus on a specialized “flip” maneuver. As flies perform it, the flip comprises a fighter-plane-like zoom into a brief vertical posture, and a further 90-degree turn to fully inverted before soft impact and perching. Visual cues are key, said Lee.
“Through video analysis of fly behavior,” said Lee, “We discovered that such aspects of inverted landing as the timing, the direction and the magnitude of the flip maneuvers are strongly correlated to the optical flows perceived by the flies.”
Cheng and Habas discerned two visual cues to creating a universal policy for drone perching: Relative Retinal Expansion Velocity (RREV) and translational optical flow.
The former is a representation of how rapidly objects expand within the camera’s field of view as the robot approaches an object, said Cheng.
“It measures how quickly, for example, the ceiling expands relative to the fly approaching it,” he explained, adding that it also happens to approximate the inverse of the time remaining for the robot to contact the surface.
Translational optical flow measures how quickly the ceiling moves in a particular direction relative to the fly, a measure that also depends on the distance between the ceiling and the fly.
“You can think of this the same way as you might think of how fast trees pass you when you are looking through a window in a fast-moving train: trees at a distance move slower,” said Lee. “It is essentially the angular velocity of the ceiling relative to the flies.”
These two inputs, RREV and translational optical flow, constitute the basis of a two-stage, general control policy for inverted landing:
“It is really a split of a second decision — a delayed trigger would cause a collision, and an early trigger would cause the robot to miss the ceiling and fall,” Lee explained.
According to Lee, the goal of these algorithmic policies, facilitated by reinforcement learning, is to enable landing not only in controlled lab environments, but in diverse, risky, unpredictable real-world conditions that might include varying velocities, different approaches and directions, obstacles and wind speed.
“We haven’t solved the entire problem yet, but focused on generalizing the control policy to arbitrary approach conditions, but not yet for obstacles, wind speed or surface orientation,” said Lee, adding that, in subsequent work, the team generalized the control policy to surfaces of arbitrary orientations.
The team is developing a robust, adaptable landing gear and optical flow estimation algorithm to enable dynamic perching under less-than-ideal conditions, such as on building walls, moving cars or tree trunks, according to Cheng.
As part of its setup, the investigators also employed a tool of their design: a Geometric Tracking Controller, tuned to such tasks as high-speed flight regulation, trajectory tracking, and updating of sensory state estimates.
Drones that can perch inverted or cling to surfaces in which gravity hinders, rather than helps, enjoy a “leg up” – so to speak – on traditional drones when it comes to rescue, construction, mapping, extraterrestrial labors, and other work in harsh conditions. In particular, Lee predicts that autonomous, universal perching ability will supercharge reconnaissance, inspection, surveillance, environmental monitoring and search and rescue opps.
“For example,” he said, “This capability will enable drones to land on, say, a sailing ship that heaves and sways with the sea, or to hitchhike onto a moving truck, or to assist a human-pilot to easily land a drone on self-selected targets, such as on walls, underneath a bridge, crops, or even a moving target.”
Added Tognon, “In energy networks, such drones could land on power lines or pylons to recharge and act as distributed sensing nodes, while in industrial or post-disaster scenarios they could cling to ceilings to provide lighting, communication, or environmental monitoring in confined or hazardous spaces.”
