Active Learning-augmented Intention-aware Obstacle Avoidance of Autonomous Surface Vehicles in High-traffic Waters

TL;DR

A topological modeling of passing intention of obstacles, which can be applied to varying encounter situations based on the inherent embedding of topological concepts in COLREGs, and a multi-objective optimization framework including information gain about the passing obstacle intention and safety are introduced.

Abstract

This paper enhances the obstacle avoidance of Autonomous Surface Vehicles (ASVs) for safe navigation in high-traffic waters with an active state estimation of obstacle's passing intention and reducing its uncertainty. We introduce a topological modeling of passing intention of obstacles, which can be applied to varying encounter situations based on the inherent embedding of topological concepts in COLREGs. With a Long Short-Term Memory (LSTM) neural network, we classify the passing intention of obstacles. Then, for determining the ASV maneuver, we propose a multi-objective optimization framework including information gain about the passing obstacle intention and safety. We validate the proposed approach under extensive Monte Carlo simulations (2,400 runs) with a varying number of obstacles, dynamic properties, encounter situations, and different behavioral patterns of obstacles (cooperative, non-cooperative). We also present the results from a real marine accident case study as well as real-world experiments of a real ASV with environmental disturbances, showing successful collision avoidance with our strategy in real-time.

Figures (8)

• Figure 1: From time $t$ to $t+1$, controlled ASV, $R$'s collision avoidance behavior by state-of-the-art method vs. proposed method using active learning-augmented intention-awareness under an uncertain scenario where an obstacle, $O$ approaches from the left side of $R$: (a) At $t$, the state-of-the-art, lacking intention-awareness, predicts that O will pass on the left side (red) of $R$ and thus $R$ maintains its course as a stand-on vessel. At $t+1$, $R$ realizes $O$ is attempting to pass on the right side (green) of $R$, resulting in a nearmiss. $R$ did a hard turn-over but it is too late. (b) At $t$, our method classifies the topological passing side based on the historical data from $t_h$ to $t$ and actively determines an action to increase information gain, i.e., to decrease the probability of passing on the right side (green), which is risky due to bow crossing. This proactive action with good seamanship despite a stand-on status leads to a safe clearance at $t+1$.
• Figure 2: Topological classification of passing and the concept of winding number changes. (a) Relative view of an example scenario where $O$ approaching from the left bow of $R$ can pass on the left ($\omega^{O|R} > 0$) with respect to $R$ (red lines) or right ($\omega^{O|R} < 0$) with respect to $R$ (green lines); and (b) action $a_t$ to $a_{t+1}$ ($\theta=90°$ to $\theta^\prime=135°$) makes $O$ pass as $\mathcal{P}_l$ by $\Delta \omega^{O|R} > 0$ with a fixed $v^R$ assumption. The state of each ship is $\mathbf{x}^R=[-30, 0], \theta^R=90°, v^R=2.5m/s$, $\mathbf{x}^O=[0, 20], \theta^O=225°, v^O=3.0m/s$, while the direction follows the maritime convention, i.e., clockwise from north.
• Figure 3: LSTM-backbone neural network structure for passing classifier.
• Figure 4: Winding number changes and expected information gain based on a next action in Fig. \ref{['fig:winding-number']} scenario. (a)$\Delta \omega^{O|R}$ under a deterministic and noisy condition with sampling $M=1,000$ with noise in Section \ref{['sec:result']}. $a_{t+1}$ has more distribution of $\Delta \omega^{O|R} > 0$, i.e., higher $p_l$ than $a_{t}$with a fixed $v$ assumption; and (b)$\Tilde{I}$ cost shows $a_{t+1}$ with better information gain (less uncertainty) than $a_t$ while $a_t$ belongs to the no-go-zone (black). Note that the best information gain occurs at $315°$ while it directs to the opposite direction to the current destination, which is handled by multi-objective optimization in this study.
• Figure 5: Comparison of trajectories in an example scenario with $30$ obstacles (cooperative obstacle: cyan, non-cooperative obstacle: gray) under ego ASV's sensible range $\mathcal{S}$$100m$ (only our method drawn in blue dots for clarity). Key obstacles within $\mathcal{S}$: $obs ~6, 14, 18$ head-on, $obs~5, 16, 25, 28$ crossing from the right, $obs~22$ anchored, $obs~12$ crossing from the left, $obs~9, 29$ overtaken. (a) After passing $obs~25$, VO$^+$, MOA, MOA$^+$ took abrupt actions with respect to $obs~16$, whereas the proposed MOA$^+$LSTM captured history of intention changes by $obs~16$ and took a smooth action to pass astern of it ($\mathcal{P}_{l}$ passing); and (b) After passing $obs~16$, VO entered between $obs~22$ and $obs~18$. However, the proposed MOA$^+$LSTM conducted a holistic consideration of $obs~18,22$ as a cluster, and took an active action by right-side turn to make both $obs~18,22$ as $\mathcal{P}_{l}$ passing. Note that there are other obstacles within $\mathcal{S}$ (e.g., $obs~7$), but our approach prioritizes obstacles based on ego-centric dynamic properties, as in Section \ref{['sec:info-gain']}.