Safety-Driven Interactive Planning for Neural Network-Based Lane Changing

Connectivity Enhanced Safe Neural Network Planner for Lane Changing in Mixed Traffic

“…Planner design in autonomous driving has been a very active research area for its potential impact in greatly improving transportation system performance and also its challenge in ensuring system safety at the same time. There is an urgent need to prevent overly conservative design while also provide safety guarantees when considering environment uncertainty and complex inter-vehicle interactions [7,19,37]. The work in [28] proposes a concept of legal safety, which means that autonomous vehicles will not be the cause of accidents and there is no collision if surrounding vehicles obey traffic rules.…”

“…At the same time, we can derive the maximum deceleration of the leading vehicle 𝐿 1 (scenario as shown in Figure 1), and then perform system analysis for the worst case and adjust trajectory to ensure safety. Here we adopt the same aggressiveness assessment method for the following vehicle 𝐹 as in [19] when vehicle 𝐹 is non-connected. For the case that the following vehicle 𝐹 is connected, we assume that it is collaborative.…”

“…As for the longitudinal motion, the optimal trajectory is analyzed intuitively when the immediate leading vehicle 𝐿 1 decelerates with 𝑎 𝑥,𝑑 [19]. It is that the ego vehicle first accelerates with 𝑎 𝑥,𝑎 and then decelerates with 𝑎 𝑥,𝑑 , and keeps the distance gap no smaller than 𝑝 𝑚 when 𝑡 ∈ [0, 𝑡 𝑦,𝑓 ].…”

“…In this section, we first demonstrate the statistical performance of our connectivity-enhanced planner framework, and then present an example with detailed trajectories. We compare system performance under a few different planners to demonstrate the effectiveness of our approach: 'CV_all' represents the complete planner framework as in Figure 2, in which we leverage information shared by all connected vehicles; 'CV_follow' means that we only use information shared by the following connected vehicle 𝐹 ; 'CV_none' means that connectivity technology is not leveraged, which is the planner in [19] and can be viewed as our main baseline; 'No_agg_assess' means that we conservatively assume following vehicle is always aggressive and disable the aggressiveness assessment function.…”

“…During the transition period to the next-generation transportation system, a mixed traffic stream of human-driven and autonomous vehicles [38], and connected and non-connected vehicles need to share the transportation network. Recent progress has been made to ensure system safety when all vehicles are of the same type [8,15], or connectivity is not enabled [17,19,28,29], and only some works [19,28] can generalize to systems with neural network based components. Lane changing, especially in congested traffic, is a challenging scenario, in which those vehicles need to closely interact with each other [20].…”