Why thrust direction is enough for translation
A quadrotor's translational dynamics depend on attitude only through the quantity R b3, the body-fixed thrust axis expressed in the inertial frame. Rotations about that thrust axis, which is the heading, do not directly influence translation. This is the fact the paper builds on: instead of regulating the entire orientation on SO(3), the authors describe orientation solely by the thrust direction, a unit vector on the sphere S2, and treat the heading as generated by an independent controller.
That choice is not only a simplification. The paper notes that heading control authority comes from differential rotor reaction torques, which are weaker than the moments used for tilting maneuvers, so aggressive heading tracking can push actuators toward saturation and degrade translational stability. Leaving heading unconstrained removes that demand from the translation problem.
The reduced model used for control has state (p, v, R3), where R3 = R b3 is the thrust direction, and control input (T, omega1, omega2): total thrust plus the first two angular velocity components. The moment-level input is deliberately not used; the paper states that generalizing the designs to the moment level requires one additional backstepping step.
Two controllers with different tradeoffs
The first design uses dynamic extension. The thrust T is appended to the state, creating the extended state (p, v, R3, T) and the input (Tdot, omega1, omega2), so that thrust enters at the same level as the rotational inputs and the position acquires a uniform relative degree. A backstepping controller then drives the velocity error e1 = v - vd(p,t) and the acceleration error e2 = (1/m) R3 T - ad(p,v,t) to zero.
The second design avoids dynamic extension and works directly on the reduced system using the geometry of S2. Its errors are e1 = v - vd(p,t) and e2 = R3 x R3d, where R3d is the unit vector along the desired acceleration profile ad. The thrust is set to m times the norm of ad, and the angular velocity command includes a term built from the cross product structure of the two thrust directions.
The tradeoff is explicit in the paper. Dynamic extension is common in practice but can introduce undesirable delays, because the designed controller must be integrated to produce the physical inputs sent to the plant. The geometric controller avoids that integration step but requires the desired acceleration to be nonzero and the thrust direction not to be antipodal to the desired one.
Exponential convergence, with stated conditions
For the dynamic extension controller, Theorem 1 gives an explicit exponential bound: the tracking error norm satisfies ||e(t)|| <= max{sqrt(sigma), 1/sqrt(sigma)} ||e(t0)|| exp(-k(t - t0)), with k = min{k1/2, k2/2}. The proof uses a Lyapunov function combining the velocity error and the scaled acceleration error, and the bound holds along trajectories on the maximal interval of existence.
For the geometric controller, Theorem 2 gives a comparable but differently parameterized bound: ||e(t)|| <= sqrt(c2/c1) ||e(t0)|| exp(-(k/2)(t - t0)), with k = min{k1, (2 - rho) k2}, c1 = min{1/(2 sigma), 1/2}, and c2 = max{1/(2 sigma), 1/(2 - rho)}. This result requires the initial condition to lie in the region N_rho(t0) for rho in (0,2) and requires ||ad(p0, v0, t0)|| > 0.
The two bounds are not interchangeable. The dynamic extension bound depends on sigma and on k1, k2 directly; the geometric bound carries the region parameter rho through both the decay rate and the constant c2, so the size of the region in which the guarantee holds affects how fast the guarantee decays. Both theorems are stated for solutions on their maximal interval of existence, and the paper explicitly declines to claim completeness of solutions, noting that forward invariance does not require it.
Encoding safety into the vector field
Safety is handled by fixing the structure of the tracking controllers and instead designing the desired vector field vd so that it satisfies a robust control barrier function condition: the gradient of h with respect to position, dotted with vd, must be at least -alpha h(p) plus (1/epsilon) times the squared norm of that gradient. Here h defines the safe set as its zero superlevel set, alpha > 0 is a linear extended class K-infinity function, and epsilon > 0 is a robustness parameter inspired by input-to-state safety.
The intuition the paper gives is direct: if the actual velocity equaled vd exactly, the condition would imply the standard CBF condition on h. Because tracking is only exponential rather than instantaneous, the controllers must additionally dominate the transient tracking error, which is why the safety theorems impose gain conditions rather than only vector field conditions.
For the dynamic extension controller, Theorem 3 requires k1 >= max{epsilon mu, 2 alpha} and k2 >= alpha, and concludes that the set defined by h(p) - (1/mu) V >= 0 is forward invariant, so p(t) stays in the safe set for all time in the interval of existence. For the geometric controller, Theorem 4 requires k1 >= max{epsilon sigma mu, 2 alpha} and k2 >= alpha/(2 - rho), together with the initial condition lying in the safe set and satisfying the conditions of Theorem 2.
The paper summarizes the meaning of these gain conditions in two parts: initialization must be inside the safe set and where the dynamics are well-defined, and vector field tracking must be sufficiently fast relative to variations in the safe vector field, as captured by alpha.
What the flight tests actually show
Simulations first test tracking without safety. With mass m = 1 kg, initial position and velocity at zero, and an initial attitude rotated by theta0 = 175 degrees, which the paper describes as nearly inverted, the desired vector field is vd(p,t) = -lambda (p - pd(t)) + pdot_d(t) with lambda = 1, tracking the reference pd(t) = [cos 2t, sin 2t, 1]. Gains are k1 = k2 = 4 for the dynamic extension controller and k1 = 4, k2 = 2 for the geometric controller. The velocity errors converge exponentially to zero in accordance with Theorems 1 and 2 even from that large rotational error.
Simulated safety enforcement uses position constraints pm = +/- 0.8 m on the first two coordinates, encoded through a smooth combination h(p) that defines the safe set, with a smooth safety filter applied to the nominal vd. Gains are k1 = k2 = 20, with safety-filter parameters alpha = 7.5, lambda_h = 5, epsilon = 2, mu = 10 for the dynamic extension controller, and alpha = 5, lambda_h = 2, epsilon = 2, mu = 10 for the geometric controller.
The experimental validation uses a Crazyflie 2.1 indoors. For trajectory tracking, the drone starts at rest on level ground at the origin, the reference is pd(t) = [0.5 cos(beta t), 0.5 sin(beta t), 0.25] with beta = 2 pi / 5, the vector field uses lambda = 2.5, and gains are k1 = k2 = 20 for dynamic extension and k1 = k2 = 10 for geometric control. For safety enforcement in flight, the position constraints are tightened to pm = +/- 0.4 m on the first two coordinates, with CBF parameters lambda_h = 10, alpha = 5, epsilon = 1, mu = 10.
The paper's own summary of these results is that the desired safety behavior is successfully encoded into the desired vector fields and accurately tracked by the controllers. The reported evidence is the position, velocity error, thrust, and angular velocity traces over roughly 25 seconds in each flight figure, not a quantified margin of constraint satisfaction.
Where these guarantees stop
The stability and safety results are stated up to the angular velocity level. The paper notes that extending the designs to the moment level, with M as the input, requires one additional backstepping step, so the theorems as written do not cover the physical moment input directly.
Both tracking theorems hold on the maximal interval of existence and make no claim that solutions are complete. The paper states that completeness would require further assumptions on vd and on the behavior of T along trajectories, and that this analysis is omitted. The geometric controller additionally requires ||ad|| > 0 and R3 not equal to -R3d, and its guarantee is conditioned on the initial condition lying in N_rho(t0) for rho in (0,2).
The experimental evidence comes from indoor flight tests with a Crazyflie 2.1, and the safety tests enforce constraints on the first two position coordinates only. The paper does not report outdoor operation, multi-vehicle coordination, or constraint satisfaction under disturbances beyond what the robustness parameter epsilon represents in the CBF condition.
Which controller to prefer, and when
The decision between the two designs turns on a specific tradeoff rather than on one being uniformly better. Dynamic extension is the more familiar technique and its exponential bound depends directly on sigma and the gains, but it requires integrating the designed controller to obtain the physical inputs, which the paper identifies as a source of undesirable delay. The geometric controller avoids that integration and works directly on the reduced system, at the cost of requiring a nonzero desired acceleration, a non-antipodal thrust direction, and an initial condition inside the region N_rho.
The safety gain conditions sharpen the comparison. Dynamic extension needs k1 >= max{epsilon mu, 2 alpha} and k2 >= alpha; geometric control needs k1 >= max{epsilon sigma mu, 2 alpha} and k2 >= alpha/(2 - rho). The geometric condition carries both sigma and rho, so the region parameter that certifies the tracking guarantee also enters the safety gain requirement.
The most decision-relevant point in the paper is therefore not that both controllers work, but that the safety guarantee is purchased through gain conditions tied to how fast tracking is relative to the safe vector field's variation, as captured by alpha. A reader comparing these designs should look first at whether the application can tolerate the integration delay of dynamic extension, and second at whether the geometric controller's region and non-antipodal conditions can be satisfied at initialization, because those conditions determine whether the safety theorem applies at all.
Sources used in this answer
Vector Field Following for Quadrotors via Reduced Attitude Representations with Applications to Safety-Critical Control
Proposes two reduced-attitude vector field tracking controllers for quadrotors, one based on dynamic extension and one on geometric control, proves exponential convergence of the translational velocity to the desired vector field, and shows that control barrier function safety constraints encoded into that vector field can be enforced on the full system, with validation in simulations and indoor Crazyflie flight tests.