Source-linked AI summary
Deep Imitation Learning for Autonomous Driving in Generic Urban Scenarios with Enhanced Safety
Jianyu Chen, Bodi Yuan, Masayoshi Tomizuka
TL;DR
Urban autonomous-driving planning is difficult for manually designed policies and existing imitation-learning systems, especially because deep neural policies lack safety guarantees. The paper combines bird-view trajectory imitation learning with a safety controller and reports strong performance in CARLA urban scenarios, while acknowledging reliance on simulator-provided ground-truth perception.
Problem
Manual policies are difficult to design and maintain for complex urban scenarios, while existing deep imitation-learning policies have limited performance and no theoretical safety guarantee.
Method
The framework learns a trajectory-planning policy from offline expert data using bird-view representations and adds a safety controller for collision-avoidance control.
Results
The final model almost perfectly solves the evaluated intersection and roundabout scenarios and significantly outperforms other methods in a new town.
Takeaways & Limitations
The system can learn a driving policy for generic urban scenarios from offline expert data while enhancing collision-avoidance safety in CARLA.
Takeaways & Limitations
The evaluation uses simulator ground-truth object and road information, so a perception module and its effect on performance remain to be studied.
Abstract
from arXiv · showhide
The decision and planning system for autonomous driving in urban environments is hard to design. Most current methods manually design the driving policy, which can be expensive to develop and maintain at scale. Instead, with imitation learning we only need to collect data and the computer will learn and improve the driving policy automatically. However, existing imitation learning methods for autonomous driving are hardly performing well for complex urban scenarios. Moreover, the safety is not guaranteed when we use a deep neural network policy. In this paper, we proposed a framework to learn the driving policy in urban scenarios efficiently given offline connected driving data, with a safety controller incorporated to guarantee safety at test time. The experiments show that our method can achieve high performance in realistic simulations of urban driving scenarios.
I. INTRODUCTION
Urban autonomous-driving decisions are difficult because road conditions, multi-agent interactions, and traffic rules vary, while manually designed policies can be sub-optimal and costly to maintain. The paper proposes imitation learning with a bird-view representation and a safety controller for complex urban driving.
- Complex road topology, geometry, markings, multi-agent interactions, and traffic rules make urban decision and planning particularly challenging.
- Manually designed policies can be sub-optimal because accurate cost functions are difficult to design and complex multi-agent behavior models are not generally solvable.
- Model-based approaches are expensive to develop and maintain because human engineers must improve their performance.
- Imitation learning avoids manually specifying the policy or cost function by learning from expert driving data.
- Existing camera-to-control imitation-learning methods mainly handle simple tasks, require extensive data, and lack transparency.
- Deep-neural-network driving policies lack theoretical safety guarantees, making safety a crucial requirement for autonomous driving.
- The proposed framework uses bird-view inputs, future trajectories, and a safe-set-based controller to improve learning efficiency and guarantee safe control.
III. FRAMEWORK OVERVIEW
The system combines a deep imitation-learning trajectory planner with a non-learning safety and tracking controller. It uses perception, routing, and stored expert observation-action data to produce trajectories and safe vehicle commands.
- The system contains a deep imitation-learning trajectory planner and a safety and tracking controller operating in a closed-loop driving environment.
- The planner receives routing and perception information and produces driving intelligence for following routes under varied road conditions.
- The framework assumes a functioning perception module, HD map data, and routing information, while focusing on decision and planning.
- Imitation learning stores expert observation-action pairs and optimizes a policy function as a supervised learning problem.
- The framework designs the observation, action, policy, loss, and data augmentation needed to obtain a robust policy.
- The agent converts perception and routing information into a bird-view image, predicts a trajectory, and passes it to the safety and tracking controller.
- The bird-view observation combines HD map, routing, traffic-light, historical-object, and historical-ego-state information, while the action is a trajectory vector.
A. Observation-action Representation
The planner replaces high-dimensional raw sensor inputs with a concise bird-view representation containing decision-relevant map, routing, traffic-light, and historical-object information. This representation is used to predict driving actions more efficiently.
- Directly mapping raw sensor data to control is inefficient and difficult to generalize because the input is high-dimensional and varies with appearance, weather, and daytime.
- The bird-view representation discards irrelevant texture while retaining information useful for decision making and planning.
- HD-map lane markings are rendered as yellow or white polylines in the two-dimensional image.
- Routing is represented by a sequence of waypoints rendered as a thick blue polyline.
- When the traffic light is red, the route changes from blue to purple; otherwise it remains blue.
- Past detected surrounding objects are rendered as green boxes, with reduced brightness indicating earlier time steps.
5) Past Ego States:
The method represents past ego states in the bird-view input and predicts a future trajectory rather than direct control. Perturbation-based data augmentation addresses error accumulation and improves policy robustness.
- 5) Past Ego States:: Past ego states are represented as red boxes with reduced brightness in the bird-view image.
- 5) Past Ego States:: The bird-view image is 192×192 pixels, covers a (40m, 40m) field of view, and places the ego vehicle at (20m, 8m).
- 5) Past Ego States:: The policy outputs a future trajectory vector instead of direct control, allowing a vehicle-specific tracking controller to follow it.
- 5) Past Ego States:: A CNN predicts H future trajectory points from the bird-view image through a VGGNet16-style convolutional network and fully connected layers.
- 5) Past Ego States:: The training objective minimizes displacement error between expert and predicted trajectory points.
- 5) Past Ego States:: Vanilla imitation learning can become unstable because accumulated prediction errors move the vehicle into unseen states where recovery is difficult.
- 5) Past Ego States:: Noise is periodically injected during data collection so the expert demonstrates recovery from perturbations, while noisy states are removed from the dataset.
V. SAFETY ENHANCEMENT & TRAJECTORY TRACKING CONTROL
The system tracks the learned future trajectory with longitudinal and lateral controllers. The controllers use target waypoints, target speed, and heading to calculate acceleration and steering commands.
- Trajectory Tracking Control: A target waypoint within the planned horizon guides both longitudinal and lateral tracking control.The waypoint is selected at m=5 in this paper, with 1 ≤ m ≤ H−1.
- Longitudinal Controller: The longitudinal controller computes desired acceleration by regulating the vehicle’s speed toward a target speed.PID control eliminates the speed tracking error between target and current speed.
- Safety Index: The safety-index illustration models a constraint around the surrounding vehicle while accounting for the relative speed of both vehicles.The constraint is visually similar to an ellipse around the surrounding vehicle.
- Lateral Controller: The lateral controller represents the ego heading and target-waypoint direction as normalized vectors to compute heading error.The ego heading vector uses the vehicle yaw angle, while the target vector points from the ego position to the waypoint.
B. Safety Enhancement Controller
The safety enhancement controller modifies tracking commands when necessary to keep the vehicle within a defined safe state set. It constructs control constraints for surrounding objects and selects a minimally modified safe command.
- Safety Enhancement Controller: The tracking controller’s acceleration and steering commands are modified when their original values are unsafe.The safety controller is added because the tracking commands alone do not guarantee safety.
- Safe Set Construction: The safe set method defines a control set whose admissible commands keep the ego vehicle safe when the current state is safe.The method uses a safety index over ego and surrounding-vehicle states, including position, velocity, and heading.
- Safety Index: The state constraint combines an ellipse-shaped distance around a surrounding vehicle with relative-speed information.Higher relative speed makes the state more likely to be unsafe.
- Control Constraints: For multiple surrounding objects, the controller intersects their control safe sets and obtains a convex polytope.The resulting admissible set constrains the combined control command.
- Safe Command Selection: A quadratic program maps the trajectory tracker’s command into the safe control set and returns a modified command.The objective uses a 2-by-2 weight matrix W.
- Modularity: The safety controller can be used with planners beyond the imitation-learning planner presented in the paper.It modifies upper-level planner control outputs to enhance safety.
A. Simulation Environment and Data Collection
The method is trained and evaluated in CARLA using a multi-agent urban simulation with intersections and roundabouts. Expert data consists of bird-view observations, vehicle state, and future trajectories recorded at 0.1-second intervals.
- Simulation Environment: CARLA provides the simulation platform for collecting data and evaluating the proposed autonomous-driving method.The platform simulates sensor data and detailed vehicle dynamics.
- Simulation Environment: The map covers 400m × 400m and approximately 6km of roads, including intersection and roundabout scenarios.The environment includes varied urban road layouts for training.
- Traffic Generation: 100 autonomous vehicles create a multi-agent traffic environment with route following, front-vehicle slowing, and red-light stopping.Vehicles randomly choose directions at intersections.
- Data Collection: The expert controller records bird-view images and ego states every 0.1 second during data collection.Recorded states include global position and yaw angle.
B. Bird-view Image Generation
The system renders historical vehicle states, map lane markings, and routing into a bird-view input, then evaluates trajectory prediction with data augmentation and example urban maneuvers.
- Bird-view representation: Historical vehicle states are transformed into the ego vehicle’s local coordinates, while lane markings and routing waypoints are rendered in the bird-view image.The buffer stores position, velocity, heading, and size for all vehicles; routing is rendered as a thick blue line.
- Training setup: The dataset contains 120k simulation frames, with 100k used for training and 20k for evaluation.Training uses Adam with an initial learning rate of 10^-4, followed by fine-tuning at 10^-5.
- Model comparison: Three models compare the effects of data augmentation and safety control: M0 has neither, M1 has augmentation only, and M2 has both.The open-loop evaluation reports average displacement error in Town03 and Town01.
- Data augmentation: Data augmentation improves recovery from abnormal states: M1 predicts trajectories that help the vehicle return to normal states more effectively than M0.Figure 5 illustrates the difference between trajectories predicted by models trained with and without augmentation.
- Urban driving examples: Model M1 produces reasonable trajectories for busy intersections, stopped or slow vehicles, red lights, and roundabout entry with yielding behavior.These examples correspond to left and right turns, stopping behind a vehicle, stopping at a red light, and entering a roundabout.
F. Closed Loop Evaluation
Closed-loop evaluation repeatedly feeds bird-view observations through the trajectory policy, tracking controller, and safety controller, and measures success and infractions in complex urban cases.
- Evaluation procedure: Every 0.1 seconds, the system renders a bird-view image, predicts a trajectory, and passes it through tracking and safety controllers before applying the control command.The loop continues until a terminal criterion is reached.
- Success rate: Success rate measures whether the vehicle reaches defined endpoints in signalized intersections and roundabouts with surrounding dynamic objects.The evaluation excludes simple lane-following cases because model M2 can succeed 100% there.
- Evaluation setup: Table II reports success percentages for intersection and roundabout scenarios across models M0, M1, and M2.Each scenario-model combination is evaluated over 50 trials.
- Infraction analysis: Infraction analysis measures the average distance traveled between collision or out-of-lane events.The comparison includes the three proposed models and Modular Pipeline, Conditional Imitation Learning, Reinforcement Learning, and Conditional Affordance Learning.
2) Evaluation Results:
The evaluation reports strong closed-loop performance, including improved success and infractions relative to learning-based baselines, while identifying failure cases and simulator-dependent limitations.
- Success-rate results: Data augmentation substantially improves recovery, and adding the safety controller enables M2 to almost perfectly solve the evaluated intersection and roundabout scenarios.Without augmentation and safety control, the vehicle can hardly pass the complex scenarios.
- Infraction results: M2 outperforms all learning-based methods on out-of-lane and collision metrics in both Town03 and Town01.Its performance is similar to the modular pipeline in the training condition and significantly better than other methods in the new town.
- Evaluation conditions: The training condition is more complex than the comparison setting because Town03 includes conditions absent from Town01, including curved roads and roundabouts.Town01 contains only single-lane roads with almost no curved roads and no roundabout.
- Failure cases: Failure cases include misinterpreting a yellow lane marking, leaving a lane without markings, and hitting a fence at a small roundabout.The authors suggest road-boundary information for the lane-marking failure and reinforcement learning penalties for obstacle collisions.
- Scope and limitations: The system is evaluated in CARLA using simulator ground-truth object and road information, which is unavailable directly in the real world.The authors identify a perception module and its effect on system performance as future work.