Abstract
This paper presents a novel trajectory planning method for aerial perching that enables micro aerial vehicles (MAVs) to land on moving, inclined surfaces. Unlike existing methods that rely on pre-determined terminal states and fixed durations, our planner adaptively adjusts these parameters to ensure safety and dynamic feasibility. By minimizing tangential relative speed, the planner is particularly effective for robots with low maneuverability or in space-constrained environments. We introduce a flexible transformation strategy to eliminate terminal constraints and reduce optimization variables, alongside precise SE(3) motion planning to prevent premature contact with the landing platform. Validated on a palm-sized MAV with a limited thrust-to-weight ratio of 1.7, the method generates optimal trajectories in under 20ms and replans in 2ms, demonstrating high efficiency and robustness.
Introduction
Perching is a critical capability for MAVs, allowing them to conserve energy by attaching to surfaces. However, executing a perching maneuver is a high-stakes task that requires extreme precision. Traditional planning methods often impose "hard-coded" terminal constraints—such as requiring a specific velocity or orientation at a fixed time. In real-world scenarios involving moving targets or cramped environments, these rigid constraints often lead to dynamic infeasibility. For instance, a drone with low thrust may be physically unable to reach a target at zero velocity without violating safety boundaries or motor limits.
Furthermore, many existing planners use overly conservative collision avoidance strategies, such as large bounding boxes, which prevent drones from approaching landing platforms closely enough to perch. This research addresses these challenges by developing a mathematically flexible and computationally efficient planner capable of running onboard in real-time.
Methodology
The proposed framework utilizes a joint optimization approach that treats both the trajectory shape and timing as variables. The system is built upon several key technical innovations:
1. MINCO Trajectory Class
The planner employs the Minimum Control Effort Polynomial Trajectory (MINCO) framework. This provides a smooth, linear-complexity mapping between path points and time allocations, ensuring the optimization remains computationally tractable for onboard hardware.
2. Flexible Terminal Transformation
To move beyond rigid terminal states, we introduce a transformation strategy that allows the optimizer to adaptively adjust tangential relative speed and terminal duration. Instead of forcing a specific landing velocity, the optimizer treats the minimization of tangential speed as a goal. This allows the drone to prioritize stability and obstacle avoidance in tight spaces, finding the safest possible approach path.
3. Smoothed Collision Avoidance
Approaching a platform requires the drone to avoid the surface until the final moment of contact. To handle this "mixed-integer" challenge, we designed a smoothed logistic function. This converts the binary collision constraint into a continuous penalty function, allowing the optimizer to find solutions much faster than traditional discrete-time solvers.
4. SE(3) Motion Planning via Hopf Fibration
To ensure the generated trajectories respect the full 3D physics of a quadrotor, we utilize Hopf fibration to map the drone's orientation and thrust directly into the planning space. This ensures that the planned maneuvers are always within the physical capabilities of the vehicle's actuators.
Experimental Results
The method was evaluated using a 0.36kg palm-sized quadrotor equipped with an NVIDIA Jetson Xavier NX. Despite a very low thrust-to-weight ratio (1.7), the planner successfully executed complex perching maneuvers.
Computational Efficiency
The planner demonstrated exceptional speed, generating initial trajectories within 20ms. When replanning with a "warm start," the computation time dropped to less than 2ms. In comparative benchmarks against state-of-the-art Nonlinear Programming (NLP) methods, our planner was significantly faster; for a -110° slope, it required only 17.7ms compared to the baseline's 81.5ms.
Maneuverability and Robustness
The system successfully perched on a vertical surface attached to a ground vehicle moving at 0.6 m/s. For low-thrust scenarios, the planner naturally generated a "swing-shaped" trajectory, using gravity to assist the drone in reaching the required orientation without overtaxing the motors. The robustness of the system was further proven through successful perching on various surfaces, including 45° inclines, vertical walls, and even a whiteboard held at arbitrary angles by a human.
Discussion and Future Work
While the current framework is highly effective, it faces certain limitations. The use of Hopf fibration introduces a mathematical singularity when the drone is completely inverted, currently preventing "ceiling perching." Additionally, the experimental setup relied on external motion capture for state estimation.
Future research will focus on integrating fully onboard perception for target detection and ego-motion. Furthermore, while the planner is mechanism-agnostic, future iterations could be tuned for specific perching hardware, such as mechanical grippers, by further refining terminal normal velocity constraints.
Summary updated