Robust Rendezvous Planning Under Maneuver Execution Errors

The problem of designing rendezvous guidance maneuver plan robust to thrusting errors is addressed in this paper. The aim of this paper is to develop tractable and robust guidance algorithms. Solving the rendezvous guidance problem via a direct approach leads to uncertain optimization problems while accounting for the Guidance, Navigation and Control (GNC) systems uncertainties and errors. A worst-case approach is considered in order to obtain tractable robust counterparts. The robustness certiﬁcates derived from these guidance algorithm provide the means to analyze the eﬀects of the considered errors on the rendezvous mission. Several types of missions tested in a linear environment are used to illustrate the methodology.


Robust rendezvous planning under maneuver execution errors
Christophe Louembet 1 , Denis Arzelier 2 , Georgia Deaconu 3 CNRS; LAAS; 7 avenue du colonel Roche, F-31400 Toulouse; France The problem of designing rendezvous guidance maneuver plan robust to thrusting errors is addressed in this paper. The aim of this paper is to develop tractable and robust guidance algorithms. Solving the rendezvous guidance problem via a direct approach leads to uncertain optimization problems while accounting for the Guidance, Navigation and Control (GNC) systems uncertainties and errors. A worst-case approach is considered in order to obtain tractable robust counterparts. The robustness certificates derived from these guidance algorithm provide the means to analyze the effects of the considered errors on the rendezvous mission. Several types of missions tested in a linear environment are used to illustrate the methodology. Any orbital operation involving at least two vehicles (satellite or International Space Station servicing, debris removal and formation flight) include a rendezvous process which consists in bringing the actuated vehicle (chaser) into the proximity of the passive vehicle (target) via a series of elementary phases (phasing, far range rendezvous and close range rendezvous). Each phase is composed of orbital maneuvers that must be carefully planned to reach their goal. This paper is particularly dedicated to the far range rendezvous operations, which begins when the separation between the chaser and the target is sufficiently small to allow the relative navigation between the two spacecraft.
From a mathematical point of view, the rendezvous guidance problem consists in an impulsive fuel-optimal control problem when the gas ejection thrust is assumed to be impulsive. This particular 2 optimal control problem can be solved using the so-called primer vector theory that originates from the seminal work of Lawden [1]. This theoretical corpus relies on optimality conditions derived from Pontryagin's Maximum principle. This indirect approach has been the base for numerous studies addressing the orbital rendezvous problem [1][2][3][4][5][6]. However, the algorithms from previously cited works are, in general, tailored for off-line computations and are not intended to be part of the spaceborne guidance software. In this context, numerical solutions based on direct methods [7] are particularly appealing. In fact, direct methods consists in parametrizing the original rendezvous guidance problem and converting it into a programming problem for which efficient solvers are available. These approaches are known to be particularly effective in presence of state and input constraints even if they generally do not provide the optimal solutions but only an approximation.
The application of direct methods to the rendezvous guidance problem was initiated by Waespy [8] and later developed by Robertson et al. in the late nineties [9]. In these works, the constraints are usually discretized in order to obtain a linear program. However, paper [10] recently proved , by converting the rendezvous problem in an semi-definite programming problem, that it is possible to satisfy the continuously in time without discretisation. Numerous paper addressing different rendezvous control problems, ranging from coordination between several spacecraft [11,12], to collision avoidance [13][14][15] often in conjunction with the Model Predictive Control (MPC) [11,12,16] have been published since then. If impulsive thrust and linearized dynamics are assumed in the previous cited papers [8][9][10][11][12][13][14][15][16], the direct methodology also permits to assume different type of thrust model or non linear dynamics for rendezvous problem. In the case of continuous thrust, the solution of the constrained and non linear optimal control problem has been obtained in [17] by performing a sequence of convex and linear constrained optimal control problems solved by either a transcription or a shooting method. The advantage of such a method is that the model can be augmented easily with orbital disturbances without increasing its complexity. In [18], the thrust is modelled by pulse and the optimal control is obtained by modulating the amplitude of the pulse addressing the fact that the thrust time can be of few minutes width. In [19], the pulse width modulation is considered and the optimal control problem is transformed in a 2-step linear program.
Among the challenges that arise from the needs of more spacecraft autonomy, the robustness 3 of the control method to the uncertainties inherent to GNC systems is a key issue [20]. These uncertainties may originate from errors in the navigation system and/or the propulsion devices. For example, in the far-range rendezvous phase, the chaser relative position and velocity provided by the navigation system may be corrupted by measurement noise. In addition, the thrusts delivered by the propulsion devices may be inaccurately executed. These errors usually lead, in real implementation, to the fact that open-loop maneuvers might steer the system far from the rendezvous objectives.
In the case of infinite control horizon, closed-loop schemes may help to mitigate GNC system's errors and ensure system stability. In the literature, numerous works propose solutions to reject perturbations and disturbances of various types. The case of impulsive rendezvous has been addressed using the discrete linear quadratic regulator (LQR) techniques [21] (see the review paper [22] and references therein). Adaptive control techniques compose a family of control laws that enables the rejection of propulsion errors in the context of formation flight by exploiting Lyapunov stability criteria as in [23]. The receding horizon MPC techniques (see, for instance, [24,26,27]) have also been successful to tackle the problem of thrust errors mitigation. However, when it comes to the finite horizon rendezvous problem, it is much more complicated to get equivalent closed-loop certificates of stability and performances (in terms of precision) using these previously mentioned techniques.
In [25], a tube-MPC scheme has been tailored to compute feedback control laws robust to navigation uncertainties while guaranteeing a given precision at the end of the rendezvous mission. Note that the errors addressed in [25] are different in nature from the propulsion errors tackled in this paper.
Indeed, navigation uncertainties are usually treated, in the literature, as additive disturbances while thrust errors are addressed in the sequel as multiplicative disturbances (see [23]). At the best of our knowledge, the generalization of tube-MPC schemes to cope with multiplicative uncertainties like the ones affecting propulsion system is a very difficult challenge.
In this paper, the rendezvous planning problem is solved using linear programming problems arising from the application of a direct methodology. In order to compute a robust maneuver plan that provides a given level of rendezvous precision, the idea is to define an a priori model of the uncertainties and errors affecting some data of the program. It is therefore necessary to resort to specific techniques for solving robust optimization problems. The usual way to address an uncertain program is to introduce in the program the probabilistic description of the uncertain data, considering them as random values. Then, the techniques from the stochastic programming theory can be applied [29]. This has been done in numerous papers in the case of continuous thrust rendezvous (see [30] and references therein), but few are those that treat the impulsive case. For instance, Luo et al. modelled the impulse execution error as a zero-mean white noise process, leading to a multiobjective program that is solved using genetic algorithms with no polynomial time convergence [31].
Other than the numerical complexity, this type of probabilistic approach can only provide probabilistic guarantees for the constraints satisfaction.
The other way to deal with uncertain program consists in a deterministic worst-case approach. Several works in the literature have addressed the robustness of the maneuver plan in the context of navigation errors. How and Tillerson described in [28] a "multiple-model" robust approach, that consists in designing the input sequence to simultaneously satisfy the constraints for several initial conditions. It is stated in [11] that testing between 1 and 10 initial conditions can lead to satisfying results. However, the authors do not provide any certificate of robustness for this approach, they only assert that such a choice is acceptable in practice thanks to properties of the linear propagation of a convex hull. A worst case approach for dealing with navigation uncertainties which provides a guaranteed certificate is described in [15]. Mueller and Larsson modelled the errors affecting the initial state as belonging to an ellipsoid and then, presented a algorithm based on linear programming that ensure the collision avoidance on a discrete time horizon.
If stochastic programming take advantage of the fact that probability distributions governing the behavior of GNC systems are usually available, the resolution technique have a high level of numerical complexity. In addition, stochastic programming approach will provide probabilistic guarantees for the constraints satisfaction while most of constraints cannot be softened in the rendezvous guidance problem. For these reasons, a deterministic worst-case approach guaranteeing a priori robustness and with good tractability is preferred in this paper.
The contribution of the present paper is to provide robust rendezvous planning algorithms that guarantee the final rendezvous conditions in the presence of GNC uncertainties and errors.

5
Technically, the aim of the proposed guidance algorithm is to prevent the spread of these GNC systems errors by working out the direct method algorithms through robust convex optimization [33]. The main advantage is to provide the mission designer with constructive numerical algorithms built on direct optimization schemes that also give a certificate of robust optimality as a by-product.
This approach makes it possible to analyze the worst possible effects of the GNC errors on a specific mission. Note that uncertainties must be described through their bounds to match the requirement of the robust optimization contrary to the stochastic programming.
The proposed guidance algorithms are built based on a three step methodology. First, a deterministic modelling of uncertainties model is proposed. Then, the effects of these errors on the optimal rendezvous guidance algorithm are exposed. Finally, a rendezvous guidance algorithm robust to above uncertainties that computes the best guaranteed precision performance is proposed.
Please note that the scope of this paper is limited to propulsion devices errors (i.e. firing time and impulse execution errors). In fact, the proposed approach is not relevant to preclude the effects of the measurement noise. This issue is tackled separately by the authors using a different methods in the article [25]. This paper is organized as follows. Sections II provides a description of the rendezvous guidance problem as it is usually presented in the literature and proposes a polytopic relaxation of the rendezvous condition. Section A presents the general guidance problem for relaxed final conditions when uncertainties are considered along with the paradigms of the robust convex optimization.
Propulsion devices errors are treated in sections III, IV and robust guidance programs are presented.
Finally, section V presents the results obtained for several mission examples in order to validate the methodology, together with nonlinear simulations results.

II. Direct resolution methods for the rendezvous guidance problem
After presenting the relative motion model and the guidance problem, the rendezvous guidance problem is formulated as a linear program using a direct approach. If the rendezvous condition is first defined as an exact condition, it is then shown that rendezvous condition must be relaxed to handle errors and uncertainties in the rendezvous guidance algorithm.

6
A. the rendezvous guidance problem This paper concentrates on the homing phase of the spacecraft rendezvous mission, which begins when the separation between spacecraft is sufficiently small to allow the guidance of the chaser spacecraft using relative navigation. In this framework, the spacecraft relative motion can be characterized expressed in the Local-Vertical-Local-Horizontal (LVLH) frame centered at the target spacecraft position (see figure 1).

Fig. 1: Illustration of the LVLH frame
It is assumed that the chaser is moved using several ergol thrusters rigidly mounted on the vehicle frame. Under this assumption, the control can be modelled as impulsive signals that instantaneously affect the velocity component of the chaser relative state, X ∈ R 6 , in the LVLH frame.
Under the impulsive control assumption, the linearized spacecraft relative dynamics of the relative state (position and velocity) X expressed with the true anomaly of the leader spacecraft ν as the independent variable are given by where A and B are the matrices corresponding to the Tshauner-Hempel linearized and simplified dynamic equations [35]. The control vectors ∆V(ν i ) are defined in the leader LVLH frame such that ∆V(ν i ) = ∆V i δ(ν − ν i ) where the vector ∆V i belongs to R 3 and δ(ν − ν i ) denotes the Dirac impulse at time ν i .

7
The rendezvous guidance problem requires the computation of a plan of N impulsive maneuvers ..,N such that the finite-time trajectory ν → X(ν) satisfies the exact initial condition X(ν 1 ) = X 1 , the final condition X(ν f ) = X f (see Fig. 2) and the actuators constraints, while minimizing a given cost (generally the ergol consumption).

Initial state
Final state Target

Fig. 2: Deterministic rendezvous illustration
The spacecraft rendezvous guidance problem can be stated as follows: where ∆v i denote the saturation bounds for the spacecraft thrusters. Note that ∆v i depend on the impulse time, ν i and on the actuators limitation ∆V max B. Direct approach: Linear Programming Direct shooting methods have shown their relevance for solving the rendezvous guidance problem in the last decade [11,13,15,37,38]. Their usage are handy since closed form solutions of the linearized dynamics (1) exist. For instance, Yamanaka and Ankersen developed in [36] a transition matrix Φ ya that enables the propagation of the chaser spacecraft relative state starting from an initial state X 1 (ν 1 ) under the action of N impulsive controls: where ν 1 < ν 2 < · · · < ν N ≤ ν. ∆V denotes the stacked control vector and where O 3 ∈ R 3×3 is the null matrix and I 3 ∈ R 3×3 is the identity matrix.
By using the closed form solution provided by the transition matrix, the dynamic program (2) can be transformed into the following static one The previous optimal control problem can be recast into a linear program by setting the impulse ..,N a priori and electing a optimization criteria, J(·),that is linear in the decision ..,N . Usually in the rendezvous problem, the the criteria represents the fuel consumption. However, it will be showed next that, in the presence of GNC system errors, the choice of the optimization criteria is of importance and needs to be made carefully in order to include the robustness requirements. The nature of the objective and the description of the cost function are discussed further in a dedicaded section after presenting the GNC system errors and their impacts on the the rendezvous guidance problem.

C. Effects of the GNC errors on the rendezvous guidance problem
As part of the GNC system, the performances of the guidance algorithms are tightly related to the performances of the navigation devices and propulsion equipment. Errors from these devices 9 and equipment can greatly deteriorate the guidance precision performances.
The navigation errors affect the initial state X 1 used for the calculation of the guidance plan: The uncertainty set U nav is directly related to the accuracy of the sensors and to the performances the navigation filter.
The impulses execution errors are twofold: errors on the firing times, ν i , and errors on the impulses execution. The spacecraft thrusts have been modelled as pure impulsive control, but practical execution of the maneuver make the trust profile looks like a continuous functions. To account for the modelling errors of the impulsive assumption, the firing times, ν i can be considered as uncertain: The impulses ∆V i , are computed in the leader's LVLH frame, but the global effort is allocated on the different available gas thrusters. The follower's thrusters also need to be aligned before each impulse execution by means of attitude slew maneuvers. Hence, the precision of the control execution depends on the precision of this alignment. Thus, the thrust allocation and the alignment process introduce uncertainties in the impulses execution: Previously, the rendezvous guidance problem has been defined with strict initial and final objectives [see Fig. 2 and Eq. (7)]. However, when the GNC uncertainties and errors (8) (9), (10) are integrated in the trajectory propagation (4), the final relative state X(ν f ) can no longer be exactly computed; In this case , the final rendezvous condition must be relaxed to enable for some tolerances on the final objective. The rendezvous terminal condition is replaced by a set membership constraint on 10 the spacecraft final relative state for all the possible instances of the uncertainties: The set T defines on admissible tolerance around the original rendezvous objective in presence of GNC errors. The rendezvous guidance problem (7) can be amended as the following uncertain program: where the set of uncertainties, V, as the direct product of the errors from (8) (9) and (10). It should be outlined that the tolerance set, T , can interpreted as an inclusion set of all possible instance of the final state X(ν f ) under GNC device errors.
In the sequel, the proposed work is focused on the impulses execution errors i.e. uncertainties on impulses time, ν and the mis-realization of the impulses, ∆V.

D. Relaxed polytopic rendezvous condition
In the proposed approach, T is chosen to be a polytopic set defined by the matrices H T ∈ R n T ×6 and K T ∈ R n T . The spacecraft relative state at the end of the rendezvous maneuver plan, X(ν f ), is required to satisfy the following matrix inequality: The K T vector is chosen such that the final set T is centered around the original rendezvous objective X f see figure 3. Thus, the polytopic version of the uncertain optimal control problem (13) 11 is described by the following program: Initial state final arrival set Without loss of generality, the tolerance polytope T can be specified as a parallelotope with the defining matrices H T and K T given by: Specifying a final tolerance in the form (16) presents computational advantages that will be exploited in what follows. Among these advantages, it enables to reduces the numbers of the defining parameters (the vector Γ for instance). Moreover, the definition (16) of H T and K T defines the set T as the inclusion set, in the interval analysis sense, for the spread zone of the different errors.

E. The objective function J(·)
The typical objective for a rendezvous guidance algorithm is the minimization of the fuel cost of the computed plan [20]. Assuming that the chaser spacecraft is equipped with six identical thrusters rigidly mounted in its axes, the total fuel consumption cost is given by [32]: 12 This piecewise linear cost function can be linearized by introducing slack variables Z ∈ R 3N in the guidance problems (7) and (15) such that and The reader will note that: On the other hand, in presence of GNC errors, besides minimizing the fuel cost, certifying the best rendezvous precision is of great interest. This precision objective can be described in terms of the dimensions of the tolerance set T which contains all the possible final states for the considered uncertainties. By considering the tolerance vector Γ as decision variable, the dimensions of the final inclusion set can be minimized while maintaining the linear structure of the cost function. In this cas the cot function can be chosen as: By choosing the cost function (20), the objective becomes able to compute the maneuver plan, ∆V, that is the least sensitive to the considered errors. Once computed, the tolerance set T gives a certified prediction of the spread zone for all the possible trajectories. It is also interesting to notice that the tolerance set has not to be set a priori but its size is minimized during the process.
Fuel is a critical ressource for the spacecraft applications and even when it is not explicitly included in the objective function, it must be ensured that the total consumption does not exceed the allocated budget. For a given fuel budget,M ∆V , the robust plan must be such that: The cost function (20) will be used in the robust algorithm presented in the next section along with the budget constraint (21) in order to achieve the best rendezvous precision while respecting the required limitations even in presences of errors.
The aim of this work is to obtain a tractable robust counterpart of the uncertain program (13) by using robust optimization concepts (cf appendix A). The uncertainties on impulse time and the errors affecting the impulses execution are considered separately and for each case a linear program is obtained.
III. Rendezvous guidance algorithm robust to uncertainties on impulses application time A. Modeling impulses firing time uncertainties using interval analysis In this section, the maneuver plan will be affected by some uncertainties on the maneuver firing time. The uncertainties are defined as: where ν 0 i is the nominal firing time, δν i is the maximal disturbance and the perturbation vector The entries of u are supposed to be independent from each other.
Uncertainties on the impulses times have two consequences on the linearized relative motion. First, since the transition matrix Φ ya depends on the true anomaly, the propagation of the relative trajectory will be affected. Second, the bounds on the control amplitudes, ∆v, also become uncertain since they depend on the position of the target spacecraft on its orbit (cf equation (3)). The present study does not address the uncertainties on the impulses bounds as their effect can be neglected since δν i is relatively small with respect to the orbital period.
In order to build the robust counterpart that accounts for the impulses time uncertainties, it is necessary to evaluate the maximum variation of the elements of the Yamanaka-Ankersen transition matrix on a given impulse time interval. Such information can be obtained through the computation of the interval inclusion function. Let a interval, denoted [x], be a connected and finite set of R defined by: 14 The inclusion function can be computed using several methods, such as the natural inclusion function, the centered inclusion function or the Taylor inclusion function [39]. To reduce the conservatism of the inclusion function, the Centred Inclusion Function (CIF) method is chosen to estimate

B. Polytopic rendezvous robust counterpart
The final rendezvous condition is relaxed to polytopic final set to account for the presence of the impulse firing time uncertainties (see Fig. 4).
Under impulse firing time uncertainties, the polytopic rendezvous constraint (14) becomes uncertain since the matrix B is unknown and belongs to the interval matrix [B] such that  Thus, the uncertain polytopic rendezvous condition is given by where The . The uncertain set V can be then expressed as: as: which is equivalent to By observing that the slack variables Z i , such that |δΦ i B∆V i | ≤ |δΦ i B||∆V i | ≤ |δΦ i |Z i , the previous inequality can be linearized to obtain the robust program to uncertain impulse times: Note that the problem (31) has a linear structure.

IV. Rendezvous guidance algorithm robust to errors in impulses execution
A. Modeling the uncertainties The performances of the thrust devices are described here in terms of amplitude and orientation precision (see figure 5).
where the coefficient λ i represents the amplitude error and is such that and the Cardan rotation matrix M ci (ψ i , θ i , φ i ) represent the orientation error (see figure 6).  As long as the Cardan angles ψ, φ, θ remain small, the Cardan rotation matrix can be approximated by where the Cardan angles are bounded by β, the maximal error angle Let M i denote the perturbation matrix such that Note that M i is composed by four independent element. thus, four perturbation variables u ij are 17 required to describe the affine uncertainty set V. where

B. Polytopic rendezvous robust counterpart
The rendezvous mission under errors on the impulse execution with a relaxed polytopic arrival set is illustrated in Fig 7. The uncertainties on the impulses execution can be translated into uncertainties on the transition matrix by using the description of the uncertainty set (37). In this case, the uncertain polytopic rendezvous condition can written as: where and the perturbation variable u ∈ R 4N is norm-bounded such that u ∞ ≤ 1. A robust counterpart can be formulated by applying the results from [40]: Then, it comes that The robust counterpart to the relaxed guidance problem (13) under uncertainties on impulses execution (32) is finally obtained: By using the slack variables Z defined by (18) and the variables Y ∈ R 4N such that the problem (45) can be reformulated as the following linear program The optimal deterministic guidance algorithm (7) minimizing the ∆V-consumption objective (17) on one hand and the robust algorithms (31) and (45)   For the robust algorithms (31) and (45), the allocated ∆V budget, M ∆V , is calculated with respect to J * ∆V , the optimal ∆V -consumption obtained by solving (7) along with the cost function (17): The aim of comparison is twofold. First, the numerical simulation must validate that the computed inclusion set T contains the spread the of errors included in the defined uncertainty sets.
The second aim of this numerical study is to ascertain the gain of the robust open-loop approach with respect to the deterministic algorithm (7)

21
In the case of the ATV 2 (Automated Transport Vehicle) mission (figures 8), the fuel-optimal plan consists of 4 impulses out of the possible 50, the robust plans are also of 4 impulses. The research of the robustness with respect to firing location errors is achieved by shifting the two first impulses to the left. In the robust plan the second and third impulses consume most part the fuel while the first and third ones are the most demanding for the fuel-optimal plan. One can also note that the impulses are only on X lvlh -axis thrusts for the three plans.
On figures 8, one can observe that the precision of the mission is improved by the robust algorithm as expected after the computation of the optimal tolerance. The improvements are mainly on Z lvlh axis on position and X lvlh axis on velocity. The bar graphic 9 exposed the level of guaranteed tolerance and the errors spread for different budgets. Different facts can be observed from 9. First, a budget of 100% of the fuel-optimal consumption is needed to guarantee a better precision than the spread by the fuel optimal plan of the firing time errors. In fact, the level of spread of errors by the non robust plan is 10% larger than the guaranteed tolerance while the actual level of spread by the robust plan is 10% smaller. Moreover, when the budget is not restricted, it can be seen that the consumption is very similar to the minimal fuel cost J * ∆V . This fact infers that the robustness do not necessarily imply an increase of the ∆V consumption but it is also a matter of impulsive thrusts location.
the advantages of using robust algorithm are more obvious in the case of Proba 3 mission. However, for this difficult mission, improvement, with respect to the spread of the fuel-optimal plan, can be obtained with less than the fuel-optimal cost . In fact, It can be seen on figure 11 that only 90% of the nominal optimal consumption, J * ∆V , is needed to certify a much better inclusion set and to greatly reduce the spread of the final disturbed states. Note that the level line of the non robust spread does not appear on figure 11 since the level of the spread is much larger than the worst computed tolerance. This fact is symbolized by the triangle. With a budget of 200% 22 of the optimal consumption, the guaranteed tolerance dimension are improved of 35%. Beyond a ∆V budget of 200% of the optimal consumption, the guaranteed tolerance is not improved significantly even if more ∆V consumption permits to tighten more the spread of errors. For both missions, one can note that the gap between the guaranteed tolerance and the actual spread of the errors is about 10 points of percentage for budget greater than 100% of the optimal-fuel consumption.  i.e. 0.1% and orientation errors β = 1 • are considered on each executed impulsive maneuver. To simulate execution errors on each impulses, the angles ψ i , θ i , φ i and amplitude coefficient λ i are chosen randomly within their bound, then the computed ∆V i is perturbed accordingly to (32). The results only present the in-plane propagation. However, one must keep in mind that mis-execution errors are spread over the three directions.
In the case of ATV 1 mission, figures 13 show that the minimal certified tolerances ensure much better precision in position than the one can be expected with the ∆V optimal plan. Moreover, the 28 spread of the errors is much more contained with the robust plan.
In Fig. 13, the optimal 4-impulses plan takes advantage of a long coasting period between the second and third impulse to reduce consumption of the ATV 1 mission. The unrestricted-budget robust maneuver plan is composed of 4 impulses. The most consuming impulsion of this robust plan is the second one that is placed halfway to the end of the mission. When limiting the budget, the robust strategy is slightly different: the halfway impulse is replaced by two impulses on X lvlh axis. However, it can be noticed that, when the robustness is sought, it is preferable to to thrust at midway to the end of the mission than having a coasting period.
Some information can be deduced from figure 14. First, the level of guaranteed tolerance is larger than the non robust plan spread of the errors, some improvement can be certified on given directions (for instance for the position). However, the spread is significantly smaller for robust plan than for optimal-fuel plan, even if it can not be certified for this particular mission. This fact is explained by the conservatism observed on figures 13, can also be highlighted in bar graph 14 since for every budget, there is a large gap between the guaranteed tolerance and the errors spread. This gap is between 65 and 35 points of percentage for budget greater than 100% of the fuel-optimal consumption. Figures 15 and 16 present more spectacular results for the Proba 3 mission since the fuel-optimal plan spread the impulses execution errors over nearly 30km along X lvlh axis while the the robust plan certify tolerance width of 750m and limit the actual spread of the errors to about 320m on the same axis. The same phenomenon can be observed on the velocity where the fuel-optimal plan spread over 17m/s along the Z axis while the certified tolerance is about 0.5m/s and the errors are spread on 20 mm/s. Observing figures 15, one can note that specific thrust instants are to be preferred when robustness is sought for this particular mission. In addition, these robust thrust locations are independent from the allocated budget.
Reminding that the triangle in figure 16 means that the non robust spread of much larger than the certified tolerances, the robust plans permits a significant reduction of the maneuver plan sensitivity with respect to impulse execution errors. However, the dimension of the minimal 29 tolerance dimension shows large conservatism. The gap between the level guaranteed tolerance and the actual spread is of between 90 and 40 % for the presented budgets.

C. Results analysis
The both tested algorithms show the advantage to significantly reduce the spread of the errors compared to the fuel-optimal plan. If this improvement can be certified a priori in the case of errors on firing time, this is not always the case for impulses mis-execution errors. In fact, for firing time error case, the computed tolerance sets are a pretty good inclusion set of the spread of the firing time errors by the robust plans. This fact is due to the fact that the computation of the transition inclusion matrix [Φ ya (ν f , [ν i ])] is pretty accurate using the centered inclusion function [39] for errors corresponding to 1 second (δν is about few milliradians). On the contrary, the propagation of the mis-execution errors, represented by matrix M through the transition matrices Φ ya (ν f , ν i ) imply wrapping effects. The wrapping effect corresponds here to the artificial enlargement of the inclusion that includes the image of an interval vector through matrix multiplication [39]. This effect is represented in the equation (46), where the slack variables Y represent the the inclusion set of the spread of errors. These effect imply that the robust algorithm will seek to certify the rendezvous precision for larger errors than needed.   A short mission duration has been chosen in order to be able to simulate the nonlinear dynamics and orbital disturbances (J 2 term and atmospheric drag mainly) while maintaining the difference 35 The results obtained for the guidance algorithm robust to impulse trigger time uncertainties when using the nonlinear dynamics are presented in Fig. 18. The robust certificates provided by the rendezvous algorithm are validated also in a nonlinear environment while the non robust plan transgress the certified tolerance bound. The robust plan has also the advantages of only consuming 0.25% more than the optimal plan.
It can be observed in Fig. 19 that the maneuver plan robust to execution errors respects the arrival tolerance set even in the presence of environmental perturbations, while the nonrobust plan violates the tolerance box in position. The price of the certified robustness here is a increase of 8% of the ∆V consumption.  The computed tolerance, the actual spread have been analysed with respect to ∆V consumption budget. It has been observed that minimizing the arrival tolerance set clearly changes the structure of the maneuver plans. Optimizing the guaranteed precision generally implies an increase of the ∆V cost. Most results show that, for a same level of consumption, a robust structure permits to mitigate the spread of the errors in comparison with the non robust optimal plan. Moreover, in some cases, the robust rendezvous algorithms are able to certify a better precision with a ∆V budget equivalent to the one obtained with the nominal plan.
A limitation of the proposed method comes from the fact that the affine description of the uncertainty sets, which the algorithms rely on, are obtained using a rather crude interval analysis, In optimization framework, an uncertain program (A1) can be defined as follows [40]: where x ∈ R n denotes the vector of decision variables. The structure of the program is supposed to be fixed and described by the cost function f 0 and the constraints applications f i . The data of the problem, Θ ∈ R m is uncertain and depends on a disturbance variable u such that each entry Θ i evolve in uncertainty set V i : The set V is the image of the disturbance set U through application Π(·). The paradigm of the robust convex optimization is based on the following statements [33]. First, x must be obtained by solving (A1) without the exact knowledge of the data Θ i . Second, the results are valid for any Thus, the optimal solution of the uncertain problem (A1) is obtained by solving a min-max problem, the so-called robust counterpart: Robust convex programming aims at achieving a tractable description of the robust counterpart (A4) [33]. In the specific case of linear programming, Thus, the nature of V is completely described by the geometry of the U set. Convex robust counterparts have been developed for specific disturbance sets U . One case of particular interest for this work is the interval case, u ∞ ≤ 1 for which the robust counterpart is also a linear program [34]. To illustrate the notion of inclusion function, consider a function f from R 2 to R 2 , with variables