TL;DR
This paper introduces a real-time spatio-temporal trajectory planning framework that enables micro aerial vehicles (MAVs) to perch on moving, inclined surfaces. By adaptively adjusting terminal states and flight duration through a novel flexible transformation strategy, the planner ensures safe, dynamically feasible perching even for drones with very low thrust-to-weight ratios.
What Problem Does This Paper Address?
Perching—the act of a drone attaching itself to a surface to save energy—is a high-stakes maneuver. Most existing methods rely on "hard-coded" terminal states, such as requiring the drone to reach a specific velocity or orientation at a fixed time.
However, in real-world scenarios with moving targets or cramped spaces, these rigid constraints often lead to dynamic infeasibility. For example, a drone with low thrust might not be able to "brake" fast enough to hit a target at zero velocity without crashing into the ground first. Furthermore, many planners simplify collision avoidance by using large bounding boxes, which makes the drone too "timid" to approach a landing platform closely. This paper aims to solve these issues by creating a planner that is both mathematically flexible and computationally efficient enough to run onboard in milliseconds.
Methods
The authors propose a joint optimization approach that treats the trajectory's shape and timing as variables. Key technical components include:
- MINCO Trajectory Class: They utilize the Minimum Control Effort Polynomial Trajectory ($\text{MINCO}$) framework, which provides a smooth, linear-complexity map between path points and time allocations.
- Flexible Terminal Transformation: Instead of fixing the terminal velocity, the authors introduce a transformation that allows the optimizer to adjust the tangential relative speed and terminal duration adaptively. This allows the drone to "find" the safest way to hit the target.
- Smooth Collision Avoidance: To handle the "mixed-integer" problem of approaching a platform (where you must avoid the platform until the very last second), they design a smoothed logistic function. This allows the optimizer to treat the collision constraint as a continuous penalty function rather than a binary "hit/miss" logic.
- SE(3) Motion Planning: By using Hopf fibration, the authors map the drone's orientation and thrust directly into the planning space, ensuring the generated trajectories respect the full 3D physics of the quadrotor.
- Optimization Solver: The problem is formulated as an unconstrained nonlinear optimization and solved efficiently using the L-BFGS algorithm.
Datasets
As a robotics systems paper, it does not use standard vision datasets. Instead, the method was evaluated through:
- Real-world Experiments: Conducted using a 0.36kg palm-sized quadrotor equipped with an NVIDIA Jetson Xavier NX.
- Comparative Benchmarking: The proposed method was compared against a state-of-the-art discrete-time Nonlinear Programming (NLP) method (Paneque et al., 2022) using simulated perching tasks on various slopes.
Key Insights and Ideas
- Adaptivity is Safety: The "aha moment" is that fixing the landing velocity is often what causes crashes. By letting the optimizer minimize tangential speed as a goal rather than a hard constraint, the drone can prioritize staying airborne and avoiding obstacles in tight spaces.
- The "Swing" Trajectory: For drones with low thrust-to-weight ratios (1.7 in this study), the planner naturally generates a "swing-shaped" trajectory. This maneuver uses gravity to help the drone gain the necessary orientation for perching without overtaxing its limited motors.
- Mathematical Smoothing: Converting complex "if-then" safety constraints into smoothed penalty functions allows high-level optimizers to find solutions much faster than traditional discrete-time solvers, which is critical for 10Hz replanning.
Key Results
- Speed: The initial trajectory is generated within 20ms, and replanning (with a warm start) takes less than 2ms.
- Efficiency: Compared to the baseline NLP method, the proposed planner was significantly faster—for a -110° slope, it took ~17.7ms vs. the baseline's 81.5ms.
- Feasibility: The system successfully perched a low-thrust MAV on a vertical surface of a ground vehicle moving at 0.6 m/s.
- Robustness: The planner handled various surfaces, including a 45° incline, a vertical wall, and a whiteboard held at an arbitrary orientation by a person.
Limitations and Future Work
- Singularity Issues: The use of Hopf fibration introduces a mathematical singularity when the drone is completely upside down. This prevents the current method from being used for "ceiling perching" or extreme inverted maneuvers.
- Sensing Dependency: The current experimental setup relies on an external VICON motion capture system or an EKF for state estimation. The authors plan to integrate fully onboard perception (target detection and ego-motion) in future iterations.
- Mechanism Specifics: While the planner is "mechanism-agnostic," the experiments used a specific magnet/adhesive hybrid. Different perching mechanisms (like grippers) might require further tuning of the terminal normal velocity.
Summary updated