Parafoil Trajectory Planning

• Parafoil trajectory planning is a control framework that guides unpowered airfoils using dynamic models for precision, safety, and energy efficiency.
• It employs methodologies such as successive convexification, NMPC, and sampling-based planners to handle non-linear dynamics and stochastic wind conditions.
• The framework integrates coordinated multi-vehicle landing, collision avoidance, and real-time replanning to ensure optimal performance in dynamic environments.

Parafoil trajectory planning encompasses the algorithmic and control frameworks for steering unpowered, gliding airfoils—principally for precision landings with rigorous safety, energy efficiency, and dynamic feasibility constraints. Core applications include autonomous supply delivery, human piloted skydiving, and large-scale coordinated multi-vehicle operations. Recent research articulates a suite of models and optimization approaches, addressing individual precision landings, human-in-the-loop training, and coordinated multi-parafoil scenarios, systematically balancing control effort, landing accuracy, wind adaptation, collision-avoidance, and online computational tractability.

1. Parafoil Dynamics and System Models

Parafoil trajectory planning employs both 6-DOF and reduced-order kinematic models depending on task scope. For detailed guidance and tracking, the full 6-DOF dynamic state for parafoil $i$ is

$$x^i(t) = \begin{bmatrix} r_R^i(t) & o_{B\leftarrow R}^i(t) & v_B^i(t) & \omega_B^i(t) \end{bmatrix}^\top \in \mathbb R^{12}, \quad u^i(t) = \begin{bmatrix} \delta_a^i(t) \ \delta_s^i(t) \end{bmatrix} \in \mathbb R^2$$

where $r_R=[x,y,z]^\top$, $o_{B\leftarrow R}=[\phi,\theta,\psi]^\top$, and $u$ consists of asymmetric/symmetric control deflections. The dynamics incorporate added-mass, aerodynamic, and wind disturbance terms as in (Wei et al., 24 May 2025).

For human-piloted and real-time planners, reduced 4-DOF kinematics are common:

$$\begin{aligned} \dot{x} &= v_h\cos\psi + w_x,\quad \dot{y} &= v_h\sin\psi,\quad \dot{z} = -v_d, \quad \dot{\psi} = u_\psi,\; |u_\psi| \le \dot\psi_{\max} \end{aligned}$$

Wind is typically altitude-dependent and modeled stochastically via Dryden turbulence filters (Leeman et al., 2021).

2. Trajectory Optimization and Control Objectives

The standard parafoil trajectory planning problem is posed as a constrained optimal-control problem:

$$\min_{u^i(\cdot), t_f^i} J = \sum_{i=1}^{N} \int_0^{t_f^i} \|u^i(t)\|^2 \,dt \;\; (+\text{terminal penalties})$$

subject to dynamic evolution $\dot x^i=f^i(x^i,u^i)$, control bounds, and terminal constraints for fixed-point precision landings:

$$x^i(t_f^i)=\theta_x^{k(i)},\; y^i(t_f^i)=\theta_y^{k(i)},\; z^i(t_f^i)=z_f,\; \psi^i(t_f^i)=\alpha_{\rm anti\_wind}$$

(Wei et al., 24 May 2025). In human-piloted scenarios, bank angle $\phi$ is the sole control input, and safety is prioritized by minimizing cumulative bank angle effort:

$J = \int_{t_0}^{t_f} \phi(t)^2 \,dt$

with final-approach constraints enforcing straight flight into the wind below a critical altitude (Fainkich et al., 11 Dec 2025).

For autonomous systems, terminal costs sharply penalize landing point and heading errors, with weights $\alpha_1 \gg \alpha_2 \gg 1$ (Leeman et al., 2021).

3. Solution Methodologies: SCP, NMPC, and Sampling-Based Planning

Key solution frameworks include:

• Successive Convexification (SCP): Non-convex control problems are iteratively relaxed to convex subproblems, each solved as a SOCP. Input substitution linearizes the state evolution, and constraints are enforced with Taylor expansions and slack variables, yielding polynomial runtime and superlinear convergence to a local optimum (Leeman et al., 2021). Real-time replanning is achieved via re-discretization and adjustment for up-to-date wind and altitude.
• Nonlinear Model Predictive Control (NMPC): For coordinated multi-parafoil scenarios, NMPC tracks reference trajectories under full nonlinear 6-DOF dynamics. At each planning instant, the controller minimizes deviation from the reference state/control sequence over a finite horizon, robustly rejecting wind and disturbances (Wei et al., 24 May 2025).
• Stable Sparse RRT (SST): Asymptotically optimal, sampling-based kinodynamic planners such as SST are adapted to account for parafoil-specific kinematic constraints and safety objectives. The search employs a bank-angle cost metric and altitude-dependent control sampling (aggressive turns at altitude, proportional wind alignment during final descent) (Fainkich et al., 11 Dec 2025). Pruning and witness sets maintain representation of locally optimal solutions.

4. Multi-Vehicle Coordination and Collision Avoidance

For multi-parafoil deployments, coordinated trajectory planning ensures energy-efficient, collision-free landings. The landing-point allocation algorithm:

1. Pre-defines $N$ candidate landing points $\{\theta^k\}_{k=1}^N$.
2. Solves/sensitizes a nonlinear program for each (parafoil, point) pair to construct the control-energy matrix $R^i(\theta^k)$.
3. Assigns each parafoil to its unique landing point via the Hungarian method for assignment minimization

$$\min_{a^{ik}\in\{0,1\}} \sum_{i,k}R^i(\theta^k)a^{ik},\qquad \sum_i a^{ik}=1,\;\sum_k a^{ik}=1$$

(Wei et al., 24 May 2025).

Collision avoidance is managed by freezing the trajectories of non-conflicted parafoils and sequentially re-solving endpoint penalized NLPs with quadratic safety margins

$\|r_R^i(t) - r_R^j(t)\|^2 \geq d_s^2$

Enforced separation is maintained with small endpoint relaxations.

5. Guidance Correction, Online Planning, and Terminal Accuracy

To compensate for kinematic model mismatch and improve trajectory accuracy, moving-horizon estimation (MHE) updates the guidance model by minimizing cumulative state deviations and parameter corrections at each step:

$$\min_{\lambda^{i}_{cor}} \sum_{l=1}^{N_{real}} \|\hat x_l-\hat x_{real}\|^2_{Q_{cor}}+\|\lambda_{cor}^i\|^2_{R_{cor}}$$

where $\lambda_{cor}^i$ adjust horizontal/vertical velocities (Wei et al., 24 May 2025).

Real-time replanning leverages updated altitude and wind estimates to recompute the time-of-flight and trajectory, maintaining responsiveness to environmental changes (Leeman et al., 2021). Simulation demonstrates reductions in landing-point dispersion (from ~30 m to 18 m), streamlined collision-free replanning (<0.5 s), and robust terminal separation (minimum 62.5 m) under wind (Wei et al., 24 May 2025).

6. Human-Piloted Parafoil Planning and Training Implications

Algorithmic planning frameworks outperform human pilots in optimality and safety. SST-based planners provide 20–80% improvement in bank-angle control effort compared to empirical pilot data, especially during intermediate flight phases where pilots typically employ tight spirals (Fainkich et al., 11 Dec 2025). The algorithm achieves precise altitude targeting for final approach while maintaining regulatory safety margins, yielding a median relative optimality gap (ROG) of ≈50%.

Pilot training programs can benefit from presenting algorithmically computed, three-dimensional optimal descent profiles rather than traditional heuristic guidance. This facilitates skill acquisition, reduces hazardous maneuvers near ground level, and promotes consistent terminal alignment into the wind (Fainkich et al., 11 Dec 2025).

7. Quantitative Performance and Simulation Analysis

Exhaustive simulation across frameworks reports:

• Monte Carlo testing (600 runs) with convex SCP guidance achieves mean landing error of ≈5 m and heading error <5°, reducing variance by an order of magnitude relative to legacy systems (Leeman et al., 2021).
• Multi-parafoil coordinated guidance achieves landing-point assignment in ≈2.2 ms ($N=6$), expedited collision replanning (total ≈0.47 s), and final landing dispersions ≤18.9 m (Wei et al., 24 May 2025).
• SST-based planners find feasible human-pilot trajectories in ≤40 s, attaining 100% success and bank-angle cost plateau after ~50 s of planning (Fainkich et al., 11 Dec 2025).

The collective advances in parafoil trajectory planning provide computationally efficient, safety-assured, and precision-capable solutions suitable for both autonomous and human-in-the-loop operations under dynamic environmental conditions.