2.1. Robot motion planning

For environment perception, a robot is equipped with different sensors, such as GPS and LiDAR, to identify its surrounding obstacles’ states, that is, positions and velocities. The motion task for the robot is to move from the initial position $\mathbf{p}_I = (x_{0}, y_{0})$ to the goal position $\mathbf{p}_G = (x_{g}, y_{g})$ with a prefer speed $v_{f}$ . Suppose the robot’s safe radius is $\rho$ , and the time is discretized into a set of time instants with an equal time step $\Delta t$ . At any time instant $t$ , $t\in \{0,1,2,\ldots \}$ , the state of the robot is described as $s_t = (\mathbf{p}_t, \mathbf{v}_{t-1}, \mathbf{p}_g, v_f, \rho )\in \mathbb{R}^{8}$ , where $\mathbf{p}_t = (x_{t}, y_{t})$ is the robot’s position at $t$ , and $\mathbf{v}_{t-1} = (vx_{t-1}, vy_{t-1})$ is the velocity in the time duration $[(t-1)\Delta t, t\Delta t)$ . The motion command at $t$ is $\mathbf{v}_{t}$ , that is, the velocity in the duration $[t\Delta t, (t+1)\Delta t)$ . Note that for unicycle kinematics, $\mathbf{v}_t$ can be represented by $(v, \theta )$ , that is, the speed and the orientation, resulting in $vx_t = v\cos{\theta }$ and $vy_t = v\sin{\theta }$ . The set of detected obstacles at $t$ is denoted as $\mathcal{O}_t =\{o_1, \ldots, o_{m_t}\}$ , and the state of each obstacle $o_i$ is denoted as $s^i_t = (\mathbf{p}^i_{t}, \mathbf{v}^i_t, \rho _i)$ , where $\mathbf{p}^i_{t} = (x^{i}_t, y_t^i)$ , $\mathbf{v}^i_t = (vx^i_t,vy^i_t)$ , and $\rho _i$ are the position, velocity, and safe radius of $o_i$ at $t$ , respectively. The state sequence of the obstacles at $t$ is denoted as $s(\mathcal{O}_t)$ . Hence, the motion problem for the robot can be described as:

(1) \begin{eqnarray} & \mathop{\arg \min }\limits _{\mathbf{v}_0, \mathbf{v}_1, \ldots, \mathbf{v}_{T-1}} T \end{eqnarray}

(2) \begin{eqnarray} s.t.\quad & \mathbf{p}_{t+1} = \mathbf{p}_t + \mathbf{v}_t \Delta t, \forall t\in \{0, 1, \ldots, T-1\}; \end{eqnarray}

(3) \begin{eqnarray} & \|\mathbf{p}_t - \mathbf{p}^i_t \| \geq \rho +\rho _i, \forall o_i \in \mathcal{O}_t, t\in \{0, 1, \ldots, T\}; \end{eqnarray}

(4) \begin{eqnarray} & \mathbf{p}_0 = \mathbf{p}_I, \ \mathbf{p}_T = \mathbf{p}_G. \end{eqnarray}

According to [Reference Chen, Liu, Everett and How12, Reference Everett, Chen and How14], such a problem can be resolved efficiently with the DRL framework by maximizing the value function:

\begin{eqnarray*} \mathbf{v}^*_t =\arg \max \limits _{\mathbf{v}\in \mathcal{A} } R(s_t, s(\mathcal{O}_t), \mathbf{v})+\gamma ^{\Delta t v_{f}} V^*(s_{t+1,\mathbf{v}}, s(\mathcal{O}_{t+1})),\\[5pt] V^*(s_{t}, s(\mathcal{O}_{t})) = \sum _{t'=t}^{T-1} \gamma ^{(t'-t)\Delta t v_f} R((s_{t'}, s(\mathcal{O}_{t'})),\mathbf{v}^*_{t'}). \end{eqnarray*}

where $\mathcal{A}$ is the set of predefined actions, $\gamma \in [0, 1]$ is a discount factor, $R(s_t, s(\mathcal{O}_t), \mathbf{v}_t)$ is one-step reward at $s_t$ by taking $\mathbf{v}_t$ , and $s_{t+1,\mathbf{v}}$ is the robot’s next state under the action $\mathbf{v}$ . Following [Reference Everett, Chen and How14], the reward function is defined as follows:

\begin{equation*} R(s_t,s(\mathcal {O}_t), \mathbf {v}_t) =\left \{ \begin {array}{ll} 1, & \mathbf {p}_{t+1} = \mathbf {p}_g \\[5pt] -0.25, & d_{\min } \leq 0\\[5pt] -0.1 + 0.5 d_{\min }, & 0 \lt d_{\min } \leq 0.2\\[5pt] 0, & otherwise \end {array} \right . \end{equation*}

where $d_{\min }$ is the minimal distance between the robot and the obstacles in $\mathcal{O}_t$ during the time duration $[t\Delta t, (t+1) \Delta t]$ .

2.2. Obstacle localization attacks (OLAs)

2.2.1. Threat model

In this paper, we consider OLAs against a robot. Particularly, as shown in Fig. 1, a robot has two ways to retrieve the surrounding obstacles’ positions. The first one is to retrieve the obstacles’ positions via the equipped sensors, such as cameras and LiDARs. The second one is via communication: When a robot can communicate with the surrounding obstacles, such as other robots in a multi-robot system, it can also retrieve the obstacles’ positions via communication. We assume that the retrieved positions may be malicious and can be modified by the adversary. Several reasons make this assumption realistic. (1) The equipped sensors may suffer from attacks, such as jamming attacks and spoofing attacks, such that they cannot be used to localize the positions of obstacles. (2) The communication network among robots is vulnerable to communication attacks [Reference Yaacoub, Noura, Salman and Chehab40, Reference Zhou, Tzoumas, Pappas and Tokekar41], and the attacker can modify data-in-transit or even cut off communication among robots [Reference Zhou and Kumar42]. (3) Some robots may be intruded into via the system vulnerabilities of the robot and send malicious messages to others [Reference Deng, Zhou, Xu, Zhang and Liu26].

Figure 1.

The adversary performs obstacle localization attacks in a robot system via different attack ways.

We assume that a robot can be under attack at some time instant while attack-free at another time instant. Hence, as shown in Fig. 1, we assume that during the motion of robots, the attacker has one of the following three capabilities: (1) interfering with the related sensors (e.g., LiDAR) such that they cannot localize the surrounding obstacles accurately, (2) modifying the positions transmitted in the communication network, and (3) manipulating a robot to send a fake position to the communication network. We also assume the attacker cannot access the control software and hence cannot bypass ObsGAN-DRL. This assumption can be guaranteed via different technologies, such as partitioning operating systems [Reference Zhao, Sanán, Zhang and Liu43]. In addition, all robots are assumed to be able to locate their own positions accurately. Several technologies can guarantee this assumption, such as multiple sensor fusion [Reference Han, Zhan, Zhao, Pan, Zhang, Wang and Liu44].

2.2.2. Example of OLAs

Under OLAs, a robot may receive fake positions of the obstacles. As shown in Fig. 2, $o_1$ and $o_2$ are two robots that can communicate with the robot $r$ . At the current time $t$ , the real position of $o_2$ is $\textbf{p}_t^2=(x_t^2, y_t^2)$ . However, due to OLAs, the received position by $r$ is $\tilde{\textbf{p}}_t^2=(\tilde{x}_t^2, \tilde{y}_t^2)$ .

Figure 2.

Effects of position errors on the motion planning of a robot. The circles represent the real positions of the two obstacle robots. The arrows indicate their motion velocities. The red dashed circle represents $r$ ’s received position of $o_2$ .

2.3. Motivation and problem statement

In the case of OLAs, there are errors between the real positions and the retrieved positions of the surrounding obstacles. Such errors may lead the robot to make a wrong decision and cause collisions. For example, as shown in Fig. 2, the robot and the two obstacles are at $\textbf{p}_t$ , $\textbf{p}^1_t$ , and $\textbf{p}^2_t$ , respectively. Therefore, the robot is expected to plan the blue path to move to the target $\textbf{p}_G$ . However, due to OLAs, the robot $r$ received a wrong position of $o_2$ , that is, $\tilde{\textbf{p}}_t^2$ . According to the wrong position, $r$ will move directly to the target. Consequently, the robot $r$ collides with $o_2$ at the position $\textbf{p}_c^2$ .

Therefore, to generate a collision-free trajectory in an adversarial environment with OLAs, the robot needs to mitigate the detrimental effects of OLAs. Hence, the problem studied in this paper can be described as follows:

3.1. GAN-based attack mitigation module

In this section, we detail the GAN-based attack mitigation strategy. Besides the primary objective of GAN models, which is generating synthetic data by learning the underlying data distribution, our method should also ensure that the generated positions accord with the kinematic constraints. Hence, we need to modify the general loss function in GAN. To correct the robot’s attacked positions, we train a GAN model to generate the potential positions of the robot under OLAs. The main training purpose is to guide the generator in learning the real data distribution and the latent features of the real positions, so the generator can approximate the real data based on historical records. Fig. 4 shows the training process of our GAN model. Since the position of an obstacle relies on the position and velocity at the previous time instant, the training data should contain the previous state to learn the latent features of the real positions. Hence, the training data can be described as $Z=\{(\textbf{p}_{t-1}, \textbf{v}_{t-1}, \textbf{p}_t)\}$ . At each episode, we sample a subset $Z_1$ to generate a set of synthetic data $\hat{Z}_1$ for the training of the generator and a subset $Z_2$ for the training of the discriminator.

(1) Generator

The generator $G$ is a multi-layer perceptron (MLP). The input of $G$ is a faked data set $\hat{Z}_1$ from the sampled real data set $Z_1$ , where

\begin{align*} \hat{Z}_1= \{\hat{z}^i_{t}= (\textbf{p}^i_{k-1},\textbf{v}^i_{t-1}, \hat{\textbf{p}}^i_{t})| \forall z_t^i = (\textbf{p}^i_{t-1},\textbf{v}^i_{t-1},{\textbf{p}}^i_{t}) \in Z_1\} \end{align*}

where $\hat{\textbf{p}}^i_{t} = (\hat{x}^i_{t}, \hat{y}^i_{t})$ is a random value simulating the results of OLAs; $\textbf{p}^i_{t-1} = (x_{t-1}, y_{t-1})$ and $\textbf{v}^i_{t-1} = (vx_{t-1},vy_{t-1})$ are the recorded position and velocity at the previous time step. Note that during the inference stage, the position at $t-1$ is the position generated by the generator at $t-1$ , rather than the retrieved one. The output of $G$ is the potential position of an obstacle: $\tilde{\textbf{p}}^i_{t} =(\tilde{x}_{t},\tilde{y}_{t})$ . Hence, we have

(5) \begin{align} \tilde{\textbf{p}}^i_{t} = G(z_t;\; \theta _G) \end{align}

where $\theta _G$ are the parameters of $G$ that need to be learned. According to the generated positions, we can obtain a set of generated samples: $\tilde{Z}_1 = \{\tilde{z}_t^i = (\textbf{p}_{t-1}^i,\textbf{v}_{t-1}^i,\tilde{\textbf{p}}_t^i) | \tilde{\textbf{p}}_t^i=G(\hat{z}_t;\; \theta _G)\}$ .

(2) Discriminator

The discriminator $D$ is another MLP. It takes a real dataset $Z_2=\{z_t^i = (\textbf{p}^i_{t-1},\textbf{v}^i_{t-1},{\textbf{p}}^i_t)\}$ and the generated dataset $\tilde{Z}_1=\{{\tilde{z}_t^i}=(\textbf{p}_{t-1}^i,\textbf{v}_{t-1}^i,\tilde{\textbf{p}}_t^i)\}$ as inputs and generates the classification result $l_t^i=D(z;\; \theta _D)$ for each $z\in \tilde{Z}_1 \bigcup Z_2$ , where $\theta _D$ are the parameters of $D$ to be learnt.

(3) Loss Functions

As the general GAN models, the loss function for the discriminator can be written as:

\begin{equation*} Loss_D = \mathbb {E}_{{z}\in {Z}_2}[\log D({z};\; \theta _D)] + \mathbb {E}_{\tilde {z}\in \tilde {Z}_1}[\log (1-D(\tilde {z}, \theta _D)] \end{equation*}

The loss function for the generator is written as:

\begin{equation*} Loss_G = \mathbb {E}_{\tilde {z}\in \tilde {Z}_1}[\log (1-D(\tilde {z};\; \theta _D)] + \frac {\sum _{z_t^i \in Z_1} \| z_t^i - \tilde {z}_t^i\|}{|Z_1|} \end{equation*}

Note that $\tilde{z}$ is function of $\theta _G$ .

Hence, the training process is to maximize $Loss_D$ while minimizing $Loss_G$ , which is shown in Algorithm 1. It is performed by optimizing the following two optimization problems in sequence via gradient descent.

(6) \begin{equation} \min \limits _{\theta _D} -Loss_D \end{equation}

(7) \begin{equation} \min \limits _{\theta _G} Loss_G \end{equation}

Algorithm 1:

Training of GAN model.

3.2. Robust motion planning

Following the GAN-based mitigation strategy, we propose our robust motion planning method ObsGAN-DRL. The architecture of ObsGAN-DRL is shown in Fig. 5, and the detailed motion planning process with ObsGAN-DRL is given in Algorithm 2. At any time instant, suppose the detected obstacles’ states are ${S}_t = \{s_t^i=({\textbf{p}}_t^i, \textbf{v}_t^i, \rho ^i) | o_i \in \mathcal{O}_t\}$ . Under ObsGAN-DRL, ${S}_t$ is processed by the well-trained generator (Lines 5–10); then, the resulting states $\tilde{S}_t = \{\tilde{s}_t^i=(\tilde{\textbf{p}}_t^i, \textbf{v}_t^i, \rho ^i) | o_i \in \mathcal{O}_t\}$ , where $\tilde{\textbf{p}}_t^i$ is generated by the generator, are sent to the DRL model to generate the corresponding command (Lines 11–17). Note that the training of the DRL model can be referred to ref. [Reference Xu, Wu, Zhou, Hu, Ding and Liu16]. In the DRL model, the obstacles are sorted in ascending order with respect to their collision criticality based on the states $\tilde{S}_t$ ; then, the corresponding LSTM model takes the sorted obstacles’ states as input and generates a unified hidden state; finally, the MLP model in the DRL model takes the robot’s state and the hidden state as input, computes the corresponding values for all action candidates, and returns the one with the maximal value. The motion is completed if the robot reaches its destination within a given error tolerance $\varepsilon$ or a collision is detected during the movement from $\textbf{p}_t$ to $\textbf{p}_{t+1}$ (Line 21).

Algorithm 2:

Motion planning with ObsGAN-DRL.

Figure 5.

The framework of ObsGAN-DRL.

4.1. Simulation setup

For the training of the GAN model, we collect 1,055,791 samples via ORCA [Reference Van Den Berg, Guy, Lin and Manocha45]. The network architectures of the generator and the discriminator are $(6,128,64,2)$ and $(6,128,64,1)$ , respectively. We train the generator and the discriminator iteratively for 50 epochs, with a batch size of 64. Their initial learning rates are 0.001. In the training phase of the generator, to simulate OLAs, the attacked position $\hat{\textbf{p}}_t$ in each $z_t$ is generated from the Gaussian distribution $N(0, 1)$ .

For training of the DRL model, the preferred speed $v_f = 1$ , the safe radius $\rho = 0.3$ , and the action space $\mathcal{A} = \{(v_i\cos{\theta _j}, v_i\sin{\theta _j})| v_i=(e^{i/5}-1)/(e-1), \theta _j=j/8, i = 0, 1, \cdots, 5,j=0,\cdots, 7 \}$ . In each replay, each obstacle is randomly located and needs to move to the opposite location with respect to the origin of the coordinates. In addition, $\eta _{gan}=10,000$ , $\mu =1$ , $\omega =50$ , $T_m=100$ , $\Delta t = 0.25$ , and $\gamma =0.9$ . The exploration rate of $\epsilon$ -greedy policy decreases linearly from 0.5 to 0.1 in the first 4,000 episodes and stays at 0.1 for the remaining episodes. The dimension of the LSTM’s hidden state is $50$ , and it is initialized with a zero vector. The architecture of the value network is (150, 100, 100, 1). The LSTM model and value network are trained simultaneously with 10 obstacles. The number of obstacles during the testing varies from 1 to 14. During the training phase, the learning rate is 0.001. All models are implemented in PyTorch. Please refer to ref. [Reference Xu, Wu, Zhou, Hu, Ding and Liu16] for the performance of the DRL-based planner under a different number of obstacles.

To show the mitigation capacity of ObsGAN-DRL, we investigate two kinds of attacks. The first one is disturbance attacks, where the position of each obstacle is modified by adding a random value. The second kind is replacement attacks, where the positions of other robots are replaced by random values. All the random values are generated from the standard normal distribution. Moreover, the attacks are launched at the 5th time step.

For testing scenarios, we randomly generate 10 test sets, each of which contains five groups based on the number of obstacles in the test cases: {1, 2, 3, 4}, {5}, {6, 7, 8, 9}, {10}, and {11, 12, 13, 14}. Each group in each set contains 300 test cases, resulting in each test set containing 1,500 test cases. We perform six experiments on the 10 test sets:

• • $normal$ : It is the baseline where the test cases are tested without any attacks or mitigations;

• • $disturbance$ : It is an attacked situation where the test cases are executed under disturbance attacks;

• • $replacement$ : It is another attacked situation where the test cases are executed under replacement attacks;

• • $normal$ + $GAN$ : It is the situation in which each test case is pre-processed by the generator;

• • $disturbance$ + $GAN$ : It is the application of ObsGAN-DRL under disturbance attacks;

• • $replacement$ + $GAN$ : It is the application of ObsGAN-DRL under replacement attacks.

4.2. Overall performance of ObsGAN-DRL

In this section, we present the experimental results of ObsGAN-DRL. Table I illustrates the overall performance across the six experiments on the ten test sets. Notably, the last two columns provide specific insights into the quality of the generated paths. From the table, we can find that the attacks not only significantly diminish the success rate (0.954 vs. 0.619 for the disturbance attack and 0.954 vs. 0.453 for the replacement attack) but also degrade the path quality, manifested in increased motion time and reduced cumulative rewards. Therefore, an attack mitigation method is necessary for adversarial environments with OLAs.

Table I.

Overall performance analysis of our generative adversarial network (GAN)-based mitigation.

Note: The sum of success, collision, and timeout rates may not equal 1 due to rounding.

From Table I, we also note a marginal decrease in the success rates for the GAN-based scenarios. To identify the reasons, we investigate the average success rates for each test group. Recall that the test cases in each test set are grouped based on the number of obstacles present. The results are given in Fig. 6. The figure shows that in the normal experiment, the average group success rate for the group with 11 to 14 obstacles is 0.821, significantly lower than the success rates for other groups ( $\geq$ 0.96). This, in turn, causes the performance of the GAN-based method to degrade from an average of 0.955 in the first four groups to 0.668 in the group with 11 to 14 obstacles. Moreover, in the GAN-based experiments, the average group success rates after GAN mitigation in this group are 0.671 and 0.680 for disturbance and replacement attacks, respectively, while the corresponding average success rates without mitigation are 0.362 and 0.178, resulting in 85.4% and 282% improvement in the average success rate. We can find that our strategy can still mitigate attacks significantly in the group with 11–14 obstacles. The primary reason for the lower success rate than other groups after mitigation is that the current normal DRL method may not perform optimally in the scenarios with 11–14 obstacles. From the results, we can conclude that the success rate of our method is dependent on the success rate of the original motion planning methods. The results indicate that the effectiveness of a mitigation method depends on both the mitigation technique itself and the original motion planning methods. Therefore, it is essential to focus not only on attack mitigation strategies but also on the development of effective motion planning methods.

Figure 6.

Average success rates for different groups in different experiments.

Therefore, in the subsequent analysis, we focus on the first four groups, that is, test cases where the number of obstacles varies from 1 to 10. It is important to note that it does not compromise the effectiveness of our method. Table II shows the overall performance of ObsGAN-DRL on the refined test cases, that is, test cases with 1–10 obstacles, and Fig. 7 shows the corresponding average success rate in each test set (10 test sets in total). Note that $disturbance$ + $GAN$ and $normal$ + $GAN$ show very similar performance, so their success rates are almost overlapped in Fig. 7. From the results of $normal$ and $normal$ + $GAN$ , we can find that the GAN-based security module does not significantly reduce the performance of the DRL-based functionality module (the average success rates are 0.988 and 0.956, respectively, and the average motion time for a scenario is 7.77 s and 7.80 s). Compared with $disturbance$ and $disturbance$ + $GAN$ , ObsGAN-DRL improves the performance significantly under disturbance attacks, increasing the success rate by 42.75% (0.955 vs. 0.669). Similarly, ObsGAN-DRL significantly improves the performance of the DRL model under replacement attacks, from 0.573 to 0.963. Moreover, according to the average motion time, the attacks increase the motion time of success scenarios and reduce the reward significantly due to the high collision rate, while the motion time and reward are similar to the normal case after GAN-based mitigation.

Table II.

Overall performance analysis for test cases with 1–10 obstacles.

Note: The sum of success, collision, and timeout rates may not equal 1 due to rounding.

Figure 7.

Average success rates for the test cases with 1–10 obstacles in the 10 sets.

In the sequel, we show the results of an example with two obstacles. Fig. 8 shows the paths under different situations, and Fig. 9 virtualizes the actual and attacked paths of the two obstacles. Fig. 8a shows the paths of the obstacles and the robot in the normal situation, while Figs. 8b and 8c show the robot’s paths under disturbance and replacement attacks, respectively. Due to the attacks, the robot collides with the obstacles without any mitigation strategy. Figs. 8d, 8e, and 8f are the paths of the robot in normal and under disturbance and replacement attacks, respectively. From the figures, we can find that the mitigated paths can achieve a similar performance to the normal one. Note that as shown in Fig. 9, the two attacks are different.

Figure 8.

An illustrative example to show the effectiveness of ObsGAN-DRL, where the circles are the end positions of the robot and the obstacles.

Figure 9.

The attacks generated in Fig. 8.

To further validate the significant improvement of ObsGAN-DRL, the analysis of statistical significance between the situations without and with GAN mitigation is conducted. In this paper, we perform t-test for equal means. The results are shown in Table III. According to the t-test results, we can conclude that the results are significant, which means that ObsGAN-DRL can mitigate attacks rather than by chance.

Finally, we show the mitigation performance of our strategy under different attacks. In detail, we generate new attacked positions from a new Gaussian distribution $N(0,2)$ and a uniform distribution $Uniform(\!-\!4,4)$ , respectively, for the disturbance and replacement attacks. The comparison results are given in Table IV. From the results, we can find that even though the original motion planning method shows different performance under different attacks, our GAN-based mitigation strategy can guarantee that the motion planning method achieves similar success rates to the normal one.

Therefore, we can conclude that ObsGAN-DRL is a general mitigation method that can effectively mitigate different OLAs against the surrounding obstacles’ positions.

4.3. Performance comparison with other mitigation methods

Currently, there is little work on the mitigation of OLAs in terms of obstacles. Hence, in this section, we show the performance comparison with the KF [Reference Suliman, Cruceru and Moldoveanu37] method and the PF [Reference Zhang, Wang and Chen39] method. In detail, KF combines the sensor readings and the predicted positions to provide more accurate position estimations. PF uses a set of weighted particles to increase localization robustness and accuracy. We use Matlab’s System Identification Toolbox to generate the dynamics for the prediction stage of KF. For PF, we initialize 50 particles using the Monte Carlo method. Fig. 10 shows the comparison of the average success rate for three mitigation methods in each test set. From the results, we can find that the proposed GAN method outperforms the other two. Specifically, compared with KF, the average success rate of the GAN method increases from 0.694 to 0.955 for disturbance, and increases from 0.693 to 0.963 for replacement. Compared with PF, the GAN method also shows a significant performance improvement (0.596 vs 0.955 for disturbance and 0.600 vs 0.963 for replacement). Table V shows the performance of the three mitigation methods in each group in terms of the average success rate. From the table, we can find that under KF, the performance reduces significantly when the number of obstacles increases, from 0.796 to 0.625 under disturbance attacks and from 0.790 to 0.619 under replacement attacks, while our GAN method can keep a relatively stable success rate. For PF, a similar diminishing trend can be observed, decreasing from 0.663 to 0.561 under disturbance attacks and from 0.666 to 0.560 under replacement attacks when the number of obstacles increases.

Table III.

Significance test for the deep reinforcement learning (DRL) model and ObsGAN-DRL under different attacks.

Table IV.

Average success rate for test cases with 1–10 obstacles under different attack strengths.

To investigate the reason for the outperformance of the GAN model, we further calculate the mitigation error of three mitigation methods. The results show that the GAN method significantly outperforms the other two methods, achieving the lowest average mitigation errors of 0.032 and 0.022 for disturbance attack and replacement attack, respectively. The reason is that by precisely correcting the attacked positions via the well-trained generator, the positions received by the DRL model are closer to the real ones. Therefore, ObsGAN-DRL can generate better motion commands.

Figure 10.

Comparison of success rates of the generative adversarial network, Kalman filter, and particle filter mitigation strategies in each test set.

Finally, to further validate the performance of ObsGAN-DRL, we perform t-test for equal mean checking. The results are shown in Table VI, which indicates the results are significant. It means that ObsGAN-DRL can achieve better performance than KF-based and PF-based methods.

4.4. Mitigation performance with other motion planning methods

Finally, we evaluate the compatibility of the GAN-based mitigation with other motion planning algorithms. In this paper, we select ORCA [Reference Van Den Berg, Guy, Lin and Manocha45] and SARL [Reference Chen, Liu, Kreiss and Alahi15]. The former is a conventional motion planning algorithm. In ORCA, each robot takes the responsibility of avoiding pairwise collisions evenly, and then the optimal action for each agent is determined by solving a low-dimensional linear program. The latter is a DRL-based method, which contains four MLPs: the first one is an embedding model, which transfers the input state to an embedding vector; the second is an attention model, which takes the embedding vector as input and computes the attention score for each obstacle; the third one is a feature module, which generates a feature vector for each obstacle; and the last is the value network, which takes the weighted feature and the robot’s state as input to compute the values. The architecture of each model can be found in ref. [Reference Chen, Liu, Kreiss and Alahi15]. The parameters for the training are the same as those given in Section 4.1.

Fig. 11 shows the success rates of either method in each test set. For ORCA, the GAN-based mitigation strategy improves the average success rate from 0.172 to 0.908 under the disturbance attack, and from 0.15 to 0.91 under the replacement attack. For SARL, the GAN-based mitigation strategy improves the average success rate from 0.66 to 0.95 under the disturbance attack, and from 0.446 to 0.956 under the replacement attack. We can find that on one hand, learning-based methods are more robust against OLAs than conventional methods; on the other hand, our proposed mitigation method can significantly improve the success rates for both conventional and learning methods. Hence, our GAN-based mitigation strategy can be compatible with other motion planning methods.

Table V.

Comparison of average success rates in each group under different mitigation strategies.

Table VI.

Significance test for generative adversarial network (GAN), Kalman filter (KF), and particle filter (PF) mitigation strategies under different attacks.

Figure 11.

The success rates of the generative adversarial network mitigation strategy with different motion planning algorithms.

4.5. Experiments in Gazebo

We evaluate our algorithm on RotorS, a well-established and high-fidelity Gazebo simulator. Gazebo is the mainstream open-source platform that can accurately reflect the physical characteristics of real-world robots. Three AscTec Firefly drones are developed to simulate the robot and obstacles in an environment. They can communicate via the topic subscription and publication in ROS. Each drone is simulated with a ROS node deployed on Ubuntu 18.04 with ROS Melodic, to execute the motion planning algorithms and generate motion commands. The safety radius of each drone is 0.3 m, and the robot drone is required to move from $(0, -4)$ to $(0, 4)$ . Note that since the rotation of rotors will cause changes in the surrounding airflow, the safety radius is larger than the physical radius of the drones. All videos of simulations can be found at https://obsgan-drl.github.io/.

4.5.1. Experiments with disturbance attack

In this scenario, the two obstacle drones are required to move from $(3.389, 1.731)$ to $(\!-\!2.067, -0.997)$ and from $(\!-\!2.313, -3.121)$ to $(1.939, 2.319)$ , respectively. Fig. 12a shows the initial states of three drones. As shown in Fig. 12b, due to the disturbance attack, the robot drone causes a collision with obstacle 0 and falls to the ground. Fig. 12c shows the real trajectories of the three drones. As shown in Figs. 13a –13c, with the proposed GAN-based mitigation strategy, the robot drone can correct attacked positions of the surrounding obstacles and navigate itself to its target successfully.

Figure 12.

RotorS experiments with two obstacles under the disturbance attack. The robot causes a collision with the obstacle and falls completely.

Figure 13.

RotorS experiments with two obstacles under the disturbance attack and mitigation. The robot arrives at the target position successfully.

4.5.2. Experiments with replacement attack

In this scenario, the two obstacle drones are required to move from $(\!-\!3.471, -2.182)$ to $(2.651, 1.628)$ and from $(\!-\!3.510, 2.054)$ to $(2.171, -1.354)$ , respectively. Figure 14a shows the initial states of three drones. As shown in Fig 14b, due to the replacement attack, the robot drone affected the rotors of obstacle drone 0 first, and then obstacle drone 0 affected the rotors of obstacle drone 1, losing the stability of the three drones. Figure 14c shows the traveled trajectories of three drones. From Figs. 15a –15c, we can find that the GAN-based mitigation strategy can deal with the attack successfully and navigate the robot drone to its target.

Figure 14.

RotorS experiments with two obstacle robots under the replacement attack. The rotors of the robot and the obstacle cause a collision.

Figure 15.

RotorS experiments with two obstacle robots under the replacement attack and mitigation. The robot arrives at its target successfully.

4.6. Discussion

In this paper, we proposed a GAN-based strategy to deal with position attacks against the surrounding obstacles. Our simulation experiments show that the proposed method can deal with such attacks efficiently. However, it also suffers from some threats concerning the training of the GAN model. On one hand, the available attack data might be limited in terms of the types and diversity of attacks in the real world, and it is challenging to collect comprehensive datasets that cover the entire spectrum of potential attacks. They will affect the generalization of the trained GAN model. To mitigate this limitation, we focus on the final influence of attacks, that is, the value of obstacle positions, which can reduce the influence of attack types and diversity; we also implement two ways to generate attacked locations. On the other hand, our attack data are generated from a simulation environment, which introduces the sim-to-real gap. Therefore, the GAN model trained in simulations may not perform as expected when deployed in the real world. To mitigate this limitation, we generate diverse data in simulations to enhance the model’s ability to handle real-world variations.