In nature, landing is far more than simply touching the ground. Many animals can land and remain attached to surfaces that would be extremely difficult for conventional aircraft to use. Flies can attach themselves to ceilings, geckos can stabilize themselves when landing on tree trunks, and bats can settle onto cave ceilings. These abilities allow animals to move, rest and survive in environments filled with irregular surfaces.
Flying robots, however, face a very different challenge. Most drones are designed to land on relatively flat, horizontal surfaces. Landing on a vertical wall requires precise control of speed, pitch angle, forces and timing. A small error during the maneuver can cause the robot to fall or become damaged.
A new machine-learning-based approach developed by Shen and colleagues aims to change this. The researchers have developed a framework that can predict whether a flying robot equipped with spines will successfully perch on a vertical wall. More importantly, the approach can perform these predictions much faster than conventional simulations and experiments, potentially making robot design and control more efficient.
Why Wall Perching Matters for Flying Robots
The ability to perch could greatly expand what aerial robots can do. Instead of continuously flying, a drone could land on a wall, power line or other suitable structure and remain there while performing a task.
This could be particularly useful in search-and-rescue missions, environmental monitoring and infrastructure inspection. A perched drone could observe an area for long periods while consuming much less energy than a drone that must constantly remain airborne.
For example, a drone equipped with a suitable perching mechanism could attach itself to a wall or elevated structure and monitor its surroundings from a useful viewpoint. This could increase operating time while reducing the energy required for flight.
However, achieving reliable perching is difficult. The robot must approach the wall with the correct combination of velocity and orientation. Its pitch angle, lift force and contact dynamics can determine whether it attaches successfully or falls away.
The Problem With Traditional Approaches
Researchers have traditionally relied on two major methods to study perching: computer simulations and physical experiments.
Although these methods can provide valuable information, they have important limitations.
A wall-perching maneuver is a rapid contact-impact event. Accurately simulating it requires very small time steps because the robot experiences sudden changes in forces when it strikes the wall. Computational methods also have to deal with complicated contact behavior and nonlinear dynamics.
As a result, simulating one landing condition can require considerable computational resources. Testing every possible combination of approach velocity and angle through simulation would therefore be extremely inefficient.
Physical experiments present a similar problem. Researchers would need to repeatedly launch the robot toward a wall under different conditions. Such experiments require preparation, equipment and time. There is also the possibility of damaging the robot when a landing fails.
This creates a major challenge: how can researchers understand a huge range of possible landing conditions without having to simulate or physically test every single one?
Machine Learning Provides a Faster Solution
Shen and the research team addressed this problem by combining traditional physics-based modeling, experiments and machine learning.
First, they developed a knowledge-based model capable of calculating the robot's transient dynamics during a perching maneuver. The model was validated against experimental results, helping establish that it could accurately represent the robot's behavior.
The researchers then used the model to identify important factors associated with successful perching, particularly under relatively low-speed landing conditions.
Next, they constructed a flying robot equipped with spines and carried out physical perching experiments. The resulting experimental data were combined with simulation data generated by the knowledge-based model.
This produced what the researchers describe as mixed sample data—a combination of real-world experimental observations and physics-based simulation results.
Machine-learning algorithms could then learn from this combined dataset.
MLP Outperformed Other Machine-Learning Methods
The researchers compared three machine-learning approaches: Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Random Forest.
The results showed a significant difference in training efficiency.
The MLP required only about 5 seconds to train, while the SVM required approximately 180 seconds because of the additional parameter-search process. In other words, the SVM took around 36 times longer than the MLP.
Prediction accuracy was also important. When the researchers examined the decision boundaries produced by the different models, SVM and Random Forest showed larger prediction errors. Some areas predicted as successful by these methods contained conditions that actually resulted in failure during numerical calculations and experiments.
The MLP showed far fewer such contradictions.
This made MLP the strongest choice for the specific wall-perching prediction problem studied by the researchers.
From “Simulation to Learning”
One of the interesting ideas introduced by this research is the concept of “simulation to learning.”
Normally, machine learning can be viewed as “learning to simulation.” A model learns from existing data and then uses that knowledge to reproduce or predict complex behavior.
Here, the researchers followed a different path.
They first used a physics-based knowledge model to generate simulation data. They then combined those results with experimental data. Machine learning was subsequently used to learn from this mixed dataset.
The advantage is that researchers do not need to physically test every possible landing condition. Instead, carefully selected experiments and simulations can provide enough information for the machine-learning model to estimate outcomes across a much larger range of conditions.
In effect, the trained model can rapidly predict whether a particular combination of robot state and landing conditions is likely to succeed or fail.
A Faster Way to Design Perching Robots
This approach could also influence robot design.
Designing a flying robot for wall perching involves many variables, including structural configuration, approach velocity and orientation. Testing every possible design through conventional simulation and experimentation would be expensive and time-consuming.
A machine-learning model can greatly reduce this burden by quickly evaluating many possible conditions.
The researchers' results suggest that such high-speed prediction could help optimize both robot control parameters and structural parameters. Instead of spending large amounts of time testing individual configurations, engineers could use the predictive model to identify promising designs more efficiently.
This could ultimately make the development of perching aerial robots faster and less expensive.
Important Limitations Remain
Despite the promising results, the system is not yet a complete solution for real-world autonomous wall perching.
The experiments were primarily performed under relatively low-speed and controlled laboratory conditions. Strong wind or other environmental disturbances could significantly affect the small robot's flight stability and therefore reduce prediction reliability.
The current robotic grasping structure also cannot successfully attach to smooth walls. Consequently, the model cannot simply be assumed to work on every type of wall or surface.
High-speed perching presents additional technical challenges. Delays in gyroscopes and accelerometers could interfere with attitude control. Motors may also have limited response speeds, making rapid transitions from flight to perching difficult. Impact-related accelerometer noise could complicate contact detection, while prolonged motor operation could create overheating problems.
These issues will need to be addressed before the technology can operate reliably in demanding environments.
The Next Step: Closed-Loop Autonomous Perching
Another major limitation is that the current framework mainly predicts the outcome of a perching event. It does not yet use that prediction as a real-time decision-making system.
In a future version, a drone could potentially predict whether its current approach is likely to succeed before reaching the wall. If failure is predicted, the robot could alter its pitch angle or approach velocity, move away from the wall and prepare for another attempt.
Such a closed-loop system would combine prediction, decision-making and control.
That would move the technology beyond simply answering “Will this landing succeed?” toward the much more powerful question: “What should the robot do to make the landing succeed?”
A Step Toward More Capable Flying Robots
Animals have spent millions of years evolving sophisticated ways to land and attach themselves to challenging surfaces. Flying robots are now beginning to reproduce some of these capabilities through a combination of robotics, physics and artificial intelligence.
Shen and colleagues' machine-learning framework represents an important step in that direction. By combining validated physics-based simulations with experimental data, the researchers created a predictive model capable of evaluating wall-perching outcomes much more efficiently than traditional approaches.
The work demonstrates how “simulation to learning” can reduce the computational and experimental burden involved in developing complex robotic behaviors.
While significant challenges remain—particularly high-speed operation, environmental disturbances, smooth surfaces and closed-loop control—the approach could help make aerial robots more mobile, energy-efficient and capable of operating in places conventional drones cannot easily reach.
In the future, drones may not always need to land on the ground. They could fly, attach to walls, rest, monitor their surroundings and take off again—bringing robotic flight one step closer to the remarkable perching abilities already found in nature.
Reference: Shen, Y., Mao, C., Qi, Z. et al. Machine learning-based framework for wall-perching prediction of flying robot. Nat Commun 16, 11038 (2025). https://doi.org/10.1038/s41467-025-67386-0
Comments
Post a Comment