IROS2019國(guó)際學(xué)術(shù)會(huì)議論文集0640

Context and Intention Aware Planning for Urban Driving Malika Meghjani1,2, Yuanfu Luo3, Qi Heng Ho1,3, Panpan Cai3, Shashwat Verma2, Daniela Rus4, and David Hsu3 AbstractWe present a novel autonomous driving system which uses the road contextual information and intentions of other road users for urban driving. Unlike highways, urban environments require the drivers to follow traffi c signs and signals while using their best judgment for anomalous situa- tions. In such scenarios, a self-driving car needs to understand and take into account the uncertainties in the environment to plan and decide its action accordingly. Our planner models the intentions of the surrounding vehicles leveraging a neural network, and integrates the road contextual information to reduce its environment uncertainties and also speed up the decision making process. We validate our planner in simulation and in a real urban environment. Our experimental results show that integrating intention inference and road contextual information for prediction, planning and decision making help improve safety and effi ciency of our autonomous driving system. I. INTRODUCTION We propose a novel autonomous driving system which uses the road contextual information and intention of other road users to provide safe and effi cient high-level driving actions for urban driving. Urban environments pose a unique set of challenges for self-driving cars. Consider a scenario for overtaking an illegally parked car from the opposite side of a two-way street with a single lane on each side. In order to decide whether to overtake the parked car, the ego- vehicle has to understand the road contextual information such as the number of lanes, direction of the lane, lane width and distance to the nearest intersection, etc. Otherwise, it can cause severe safety hazards due to misjudgments. Another challenge in urban driving is the need for long- term planning while interacting with multiple exo-vehicles. A driving system has to perform long-horizon planning in a large state space composed of all neighboring vehicles, so that the ego-vehicle avoids collisions with them while effi ciently navigating to its goal. The key aspect, in this case, is to predict the long-term behaviours of exo-vehicles and plan for the ego-vehicle, accordingly. We developed a hierarchical prediction model for long- term planning. At the high level, we model the intentions *Both authors contributed equally. MalikaMeghjani,QiHengHoandShashwatVermaarewith theSingapore-MITAllianceforResearchandTechnology1, Singapore.Malikaisalso affi liatedwithSingaporeUniversityof TechnologyandDesign2.malika .sg, ,shashwatverma14 YuanfuLuo,PanpanCaiandDavidHsuarewith theNationalUniversityofSingapore 3, Singapore yuanfu,caipp,.sg Daniela Rus is with the Massachusetts Institute of Technology4, Cam- bridge, MA, USA Fig. 1: Intention and trajectory prediction of the exo-vehicles given their past trajectories and the road contextual infor- mation. The perception system detects the vehicle with a bounding box (left). The intention and trajectory prediction results are presented on the right in text and as green markers respectively. of exo-vehicles, such as keeping the current lane, changing to the left lane, or changing to the right lane. At the low level, we use a polynomial curve fi tting method to predict the vehicles actual motion conditioned on their intentions. This hierarchical model allows us to learn the driving intentions independently of driving trajectories, thus requiring limited amount of training data to learn the correlations between the road contextual information and drivers intentions. Since the trajectories can vary signifi cantly across different drivers, the online polynomial curve fi tting model helps us capture this variance, aptly. The planning process can be formalized as a Partially Observable Markov Decision Process (POMDP) 1 which provides a principled way to handle uncertainties such as par- tial observability, action noise and sensing noise. However, POMDP planning suffers from its well-known high computa- tional complexity. In order to achieve real-time planning, we further use road contextual information to assist the search by pruning invalid actions and shaping the rewards. Our contributions in this paper are: (a) an autonomous driving framework that can track and predict intentions and trajectories of multiple exo-vehicles, (b) a context-aware prediction model which decouples intention and trajectory prediction for exo-vehicles, (c) a context and intention aware planner that determines long-term high-level actions of the ego-vehicle under uncertainties of exo-vehicles intentions. We evaluate our driving system qualitatively and quanti- tatively in a range of scenarios using a simulator which integrates real-world road networks and the relevant contex- tual information. We also analyzed our system performance on real-world data. Our results show that integrating road contextual information and intention inference into long- term planning helps improve the effi ciency and safety of the system. 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) Macau, China, November 4-8, 2019 978-1-7281-4003-2/19/$31.00 2019 IEEE2891 II. RELATED WORK A. Road contextual information Given the widespread availability of precise and high resolution digital map information 2, 3, the possibility of using the road contextual information as a prior for behavior prediction is becoming increasingly prevalent. Specifi cally, in 4, the authors propose a one-stage detector and behavior predictor using two different 2D convolution neural net- works, each one processing 3D point cloud data and dynamic map information. Similar to our work, their map information comprises static road features such as lanes, intersections, crossing and traffi c signs. They categorize behavior predic- tion into high-level actions and motion estimation using the same neural network framework. In our work, we use road contextual information for not merely behaviour prediction but also for high-level planning. B. Intention Inference and Trajectory Prediction Intention and trajectory prediction of dynamic obstacles are two key components for decision making of autonomous vehicles. A survey of state-of-the-art intention and trajectory prediction algorithms is presented in 5. Neural networks are popularly known to be useful for intention prediction 6. For example, neural networks were used for both intention and trajectory prediction for vehicles on highways in 7 and 8, and for trajectory prediction only in 9. These approaches, however, require a huge amount of data for learning a range of driving behaviors. We overcome the challenge of large data requirements by using a neural network to only predict the intention whereas the trajectory of predicted intention is obtained based on polynomial fi tting and extrapolating the real-time vehicle state 10. This allows us to predict the intention and trajectory in real-time. C. Planning with human intentions Several previous work use POMDPs to handle the uncer- tainty in human drivers intentions. Some of this research focuses on intersection scenarios with a small number of agents. In 11, the authors modeled exo-vehicles intended behaviours: to drive aggressively or patiently, as hidden vari- ables of the POMDP. They then solved the POMDP offl ine to control ego-vehicles speed at the intersection. Due to the high computational cost of offl ine planning, the approach has only been tested in two-vehicle interactions. A similar approach was taken in 12, but they infer the intention of other vehicles using a reaction-based probabilistic model. Noticeably, the work used a rich representation of road contexts to help predict other vehicles motion. However, this representation also induces high computational complexity. Thus their results only discussed interactions among 2 3 vehicles. Our system uses a much more concise notion of road contexts. Recently, POMDP planning is applied to leverage a road network known a-priori to the robot vehicle for driving at intersections 13. Their method models the intended paths of exo-vehicles on the road network as hidden variables and also control the vehicle speed. We argue that the lane merging problem is much harder than the intersection case, because exo-vehicles have a lot more freedom: drivers can choose to merge lane at any moment when they feel it promising. The decision making of lane merging scenarios has been studied by several previous work. A simplifi ed approach 14 was proposed to make the lane merging problem tractable: evaluate a fi xed set of policies by sampling exo-vehicles intentions and rolling out future interaction trajectories. A similar multi-policy decision making approach 15 has been applied to navigation among multiple pedestrians. These multi-policy methods only plan for one-time interaction with other agents. However, in real-life driving scenarios, long-term interactions are often required, e.g., executing multiple lane merges to reach a faster lane. Another set of work 16, 17 addresses the lane changing problem from the perspective of active information gathering. Both work apply exploration bonuses on the reward function to encourage probing actions that bring information to better understand human drivers inner states. Again, these work only focused on the interactions among a small number of vehicles. Another group of work studied the interaction with multi- ple agents, typically pedestrians, which incurs exponentially higher complexities than the aforementioned scenarios. It has been proposed in 18 to model pedestrian intentions as a fi nite set of goals and apply a simple goal-directed motion model to predict pedestrian motions. The method used a state-of-the-art online POMDP planning algorithm DESPOT 19 to handle moderately dense crowds. A recent work 20 proposed a more sophisticated pedestrian motion model, PORCA, and integrated it into parallel POMDP planning 21 to drive an autonomous vehicle among many pedestri- ans. Different from these work for free-walking pedestrians, our system interacts with multiple vehicles on urban roads, in which case it is important to leverage road contexts in intention inference, motion prediction, and decision making. III. OVERVIEW The overview of our autonomous driving system is pre- sented in Fig. 2. It comprises three sub-systems: perception, high-level decision maker, and low-level controller. In the perception system 22, a LiDAR-based point cloud clus- tering module and a vision-based obstacle detection and classifi cation module are used for identifying the road region and the vehicles. Their outputs are fused in the sensor fusion module for vehicle tracking. The high-level decision maker receives the tracking trajectories, infers a belief over the intentions of each vehicle, then plans lane-keeping/lane- changing action based on the inferred intention belief and the road contextual information. The planned action can then be sent to the low-level controller to plan a path and provide the steering and throttle control that tracks the path. The low-level controller is however not addressed in this paper. This work focuses on developing the decision maker to plan for high-level action, specifi cally, the action of keeping lane, changing to left lane, or changing to right lane, for the ego-vehicle, to achieve context and intention aware planning 2892 Vision LiDAR Tracking it characterizes the imperfect robot control and environment dynamics. The observation function O(s,a,z) = p(z|a,s) represents the probability of receiving an observation z after the robot executes a and reaches s; it models the sensor noises. The reward function R(s,a) defi nes a real-value reward for executing a in s. The robot does not know the exact state it is in because of imperfect sensing. Therefore, it maintains a belief, i.e., probability distribution, over S. Initially, the robots belief is b0, and it gets updated via Bayes rule at each time step. The aim of POMDP planning is to fi nd a policy , a mapping from a belief b to an action a, that maximizes the expected total discounted rewards: V(b) = E ? X t=0 tR(st,(bt) ? ? ? b 0= b ? ,(7) where stis the state at time t, (bt) is the action that the policy chooses at time t, and (0,1) is a discount factor that places preferences for immediate rewards over future ones. 2) Context and Intention Aware POMDP: We construct the context and intention aware POMDP model for lane- keep/lane-change decision making in urban environments. State Modeling. In our problem formulation, the state is defi ned as a combination of road contextual information c, pose (x,y,) of each vehicle, and the intention g of each exo-vehicle. The intentions of exo-vehicles are not observable to the ego-vehicle; we formalize them as hidden variables in the state. Action Modeling. We plan the high-level action for the ego-vehicle at each time step, includingLANE-KEEP,LEFT- LANE-CHANGEandRIGHT-LANE-CHANGE. We further prune the forbidden actions in different lanes with the help of road contextual information. Observation Modeling. The observation consists of the road contextual information, poses, speeds and intentions of 2894 all vehicles. We do not consider sensor noises on the obser- vations, to focus on modelling the uncertainty in intentions of the exo-vehicles. Transition Modeling. The transition function models the movements of each vehicle under different intentions. For each vehicle, we use the trajectory predictor from Section IV- C to predict its next-step pose, given its intention and the road contextual information. By adding a Gaussian noise on the pose, for each vehicle, we obtain a transition function: p(xt+1,yt+1,t+1|h(t),g,c),(8) where h(t) = xt,yt,t,.,xt3,yt3,t3 is a 4-time- step history of the past poses. For the ego-vehicle, its intention is represented by its high-level action. Reward Modeling. The reward function is designed for the sake of safety, effi ciency and smoothness of driving. To achieve safety, we penalize the collision with exo-vehicles with a huge penalty R = 1000 max(4 d)2,1, where d 4 meters is the distance between the two vehicles in collision. We divide the reward for effi ciency into global reward and local reward. For global reward, we assign a reward R = 0 when the vehicle reaches its destination and a penalty R = 100 (d/dmax) where d is the distance from current lane to the destination lane and dmaxis the maximum inter-lane distance, to penalize the ego-vehicle for choosing a lane that is farther to the destination. For local reward, we assign a penalty R = 20 vvmax vmax to encourage the ego- vehicle to choose a lane on which it can drive faster. For smoothness, we assign a penalty R = 1 for doing lane changes to avoid excessive lane changes. The fi nal reward is the weighted sum of the aforementioned individual rewards. Initial Belief. We use the LSTM intention predictor from Section IV-B to infer the belief over the intentions of each exo-vehicle, and use the inferred belief as the initial belief of the POMDP model. 3) Solving POMDP: We use Determinized Sparse Par- tially Observable Trees (DESPOT) 19 for solving the lane-keeping/lane-changing POMDP. DESPOT is one of the fastest online POMDP solvers. The key idea of DESPOT is to search a belief tree under K sampled scenarios only, which greatly reduces computational complexity, making it an effi cient solver for our POMDP model. V. EXPERIMENTAL RESULTS We validate our POMDP high-level planner with both simulated and real-world data, both qualitatively and quan- titatively. Specifi cally, for qualitative analysis, we designed four scenarios to validate the behavior of our planner, demon- strating the benefi t of using the contextual information, intention inference, and long-term planning. For quantitative analysis, we randomly generated scenarios in simulation and tested the performance of our planner, and compared the average results with those of baselines. In the following, we will fi rst introduce our baseline algorithms and the criteria for our performance comparison. Then we will introduce the four scenarios we designed and analyze our results with both simulated and real-world data. A. Baseline algorithms We compared our algorithm, Context-Intention-POMDP, with three baselines: Reactive-Controller, Greedy-Controller, and SimMobilityST. 1) Reactive-Controller: This controller reacts based on the distances from the ego-vehicle to the front exo-vehicles in different lanes. It fi rst gets the headway distances from the perception system, i.e., the distances to the nearest front vehicle, in the current lane, the headway distance in the left lane (if left lane exists), and the headway distance in the right lane (if right lane exists). It then compares those headway distances with a distance threshold D. It chooses to keep lane if the headway distance of the current lane is larger than D. Otherwise, it chooses to change to the lane with the largest headway distance. We set D = 20 meters in our experiment. 2) Greedy-Controller: The controller greedily chooses to drive in the lane that is shortest to the destination lane at each time step regardless of exo-vehicles. 3) SimMobilityST: SimMobilityST is a rule-based algo- rithm used in SimMobility short-term simulator 25 to cen- trally control the driving behavior of the simulated vehicles. SimMobilityST uses a four-level decision maker to model the individual driving behavior of each vehicle. These decisions include, (a) target lane selection, (b) gap acceptance of the leading and lagging vehicles, (c) the target gap and (d) desired acceleration. The fi rst decision of target lane selection, is based on a global p