For simulation to be an effective tool for the development and testing of autonomous vehicles, the simulator must be able to produce realistic safety-critical scenarios with distribution-level accuracy. However, due to the high dimensionality of real-world driving environments and the rarity of long-tail safety-critical events, how to achieve statistical realism in simulation is a long-standing problem. In this paper, we develop NeuralNDE, a deep learning-based framework to learn multi-agent interaction behavior from vehicle trajectory data, and propose a conflict critic model and a safety mapping network to refine the generation process of safety-critical events, following real-world occurring frequencies and patterns. The results show that NeuralNDE can achieve both accurate safety-critical driving statistics (e.g., crash rate/type/severity and near-miss statistics, etc.) and normal driving statistics (e.g., vehicle speed/distance/yielding behavior distributions, etc.), as demonstrated in the simulation of urban driving environments. To the best of our knowledge, this is the first time that a simulation model can reproduce the real-world driving environment with statistical realism, particularly for safety-critical situations.
SEPARATING THE WORLD AND EGO MODELS FOR SELF-DRIVING
Training self-driving systems to be robust to the long-tail of driving scenarios is a
critical problem. Model-based approaches leverage simulation to emulate a wide
range of scenarios without putting users at risk in the real world. One promising
path to faithful simulation is to train a forward model of the world to predict
the future states of both the environment and the ego-vehicle given past states
and a sequence of actions. In this paper, we argue that it is beneficial to model
the state of the ego-vehicle, which often has simple, predictable and deterministic
behavior, separately from the rest of the environment, which is much more complex
and highly multimodal. We propose to model the ego-vehicle using a simple
and differentiable kinematic model, while training a stochastic convolutional forward
model on raster representations of the state to predict the behavior of the rest
of the environment. We explore several configurations of such decoupled models,
and evaluate their performance both with Model Predictive Control (MPC)
and direct policy learning. We test our methods on the task of highway driving
and demonstrate lower crash rates and better stability. The code is available at
https://github.com/vladisai/pytorch-PPUU/tree/ICLR2022.
more »
« less
- Award ID(s):
- 1922658
- NSF-PAR ID:
- 10351036
- Date Published:
- Journal Name:
- ICLR 2022 workshop on Generalizable Policy Learning in the Physical World. 2022.
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
more » « less
Abstract -
Multi-agent autonomous racing is a challenging problem for autonomous vehicles due to the split-second, and complex decisions that vehicles must continuously make during a race. The presence of other agents on the track requires continuous monitoring of the ego vehicle’s surroundings, and necessitates predicting the behavior of other vehicles so the ego can quickly react to a changing environment with informed decisions. In our previous work we have developed the DeepRacing AI framework for autonomous formula one racing. Our DeepRacing framework was the first implementation to use the highly photorealisitc Formula One game as a simulation testbed for autonomous racing. We have successfully demonstrated single agent high speed autonomous racing using Bezier curve trajectories. In this paper, we extend the capabilities of the DeepRacing framework towards multi-agent autonomous racing. To do so, we first develop and learn a virtual camera model from game data that the user can configure to emulate the presence of a camera sensor on the vehicle. Next we propose and train a deep recurrent neural network that can predict the future poses of opponent agents in the field of view of the virtual camera using vehicles position, velocity, and heading data with respect to the ego vehicle racecar. We demonstrate early promising results for both these contributions in the game. These added features will extend the DeepRacing framework to become more suitable for multi-agent autonomous racing algorithm developmentmore » « less
-
The goal of this study was to evaluate driver risk behavior in response to changes in their risk perception inputs, specifically focusing on the effect of augmented visual representation technologies. This experiment was conducted for the purely real-driving scenario, establishing a baseline by which future, augmented visual representation scenarios can be compared. Virtual Reality (VR), Augmented Reality (AR) and Mixed Reality (MR) simulation technologies have rapidly improved over the last three decades to where, today, they are widely used and more heavily relied upon than before, particularly in the areas of training, research, and design. The resulting utilization of these capabilities has proven simulation technologies to be a versatile and powerful tool. Virtual immersion, however, introduces a layer of abstraction and safety between the participant and the designed artifact, which includes an associated risk compensation. Quantifying and modeling the relationship between this risk compensation and levels of virtual immersion is the greater goal of this project. This study focuses on the first step, which is to determine the level of risk perception for a purely real environment for a specific man-machine system - a ground vehicle – operated in a common risk scenario – traversing a curve at high speeds. Specifically, passengers are asked to assess whether the vehicle speed within a constant-radius curve is perceived as comfortable. Due to the potential for learning effects to influence risk perception, the experiment was split into two separate protocols: the latent response protocol and the learned response protocol. The latent response protocol applied to the first exposure of an experimental condition to the subject. It consisted of having the subjects in the passenger seat assess comfort or discomfort within a vehicle that was driven around a curve at a randomlychosen value among a selection of test speeds; subjects were asked to indicate when they felt uncomfortable by pressing a brake pedal that was instrumented to alert the driver. Next, the learned response protocol assessed the subjects for repeated exposures but allowing subjects to use brake and throttle pedals to indicate if they wanted to go faster or slower; the goal was to allow subjects to iterate toward their maximum comfortable speed. These pedals were instrumented to alert the driver who responded accordingly. Both protocols were repeated for a second curve with a different radius. Questionnaires were also administered after each trial that addressed the subjective perception of risk and provided a means to substantiate the measured risk compensation behavior. The results showed that, as expected, the latent perception of risk for a passenger traversing a curve was higher than the learned perception for successive exposures to the same curve; in other words, as drivers ‘learned’ a curve, they were more comfortable with higher speeds. Both the latent and learned speeds provide a suitable metric by which to compare future replications of this experiment at different levels of virtual immersion. Correlations were found between uncomfortable subject responses and the yaw acceleration of the vehicle. Additional correlation of driver discomfort was found to occur at specific locations on the curves. The yaw acceleration is a reflection of the driver’s ability to maintain a steady steering input, whereas the location on the curve was found to correlate with variations in the lane-markings and environmental cues.more » « less
-
more » « less
Abstract Vehicle‐to‐Everything (V2X) communication has been proposed as a potential solution to improve the robustness and safety of autonomous vehicles by improving coordination and removing the barrier of non‐line‐of‐sight sensing. Cooperative Vehicle Safety (CVS) applications are tightly dependent on the reliability of the underneath data system, which can suffer from loss of information due to the inherent issues of their different components, such as sensors' failures or the poor performance of V2X technologies under dense communication channel load. Particularly, information loss affects the target classification module and, subsequently, the safety application performance. To enable reliable and robust CVS systems that mitigate the effect of information loss, a Context‐Aware Target Classification (CA‐TC) module coupled with a hybrid learning‐based predictive modeling technique for CVS systems is proposed. The CA‐TC consists of two modules: a Context‐Aware Map (CAM), and a Hybrid Gaussian Process (HGP) prediction system. Consequently, the vehicle safety applications use the information from the CA‐TC, making them more robust and reliable. The CAM leverages vehicles' path history, road geometry, tracking, and prediction; and the HGP is utilized to provide accurate vehicles' trajectory predictions to compensate for data loss (due to communication congestion) or sensor measurements' inaccuracies. Based on offline real‐world data, a finite bank of driver models that represent the joint dynamics of the vehicle and the drivers' behavior is learned. Offline training and online model updates are combined with on‐the‐fly forecasting to account for new possible driver behaviors. Finally, the framework is validated using simulation and realistic driving scenarios to confirm its potential in enhancing the robustness and reliability of CVS systems.
-
null (Ed.)Key to the effectiveness of schedule-driven approaches to real-time traffic control is an ability to accurately predict when sensed vehicles will arrive at and pass through the intersection. Prior work in schedule-driven traffic control has assumed a static vehicle arrival model. However, this static predictive model ignores the fact that the queue count and the incurred delay should vary as different partial signal timing schedules (i.e., different possible futures) are explored during the online planning process. In this paper, we propose an alternative arrival time model that incorporates queueing dynamics into this forward search process for a signal timing schedule, to more accurately capture how the intersection’s queues vary over time. As each search state is generated, an incremental queueing delay is dynamically projected for each vehicle. The resulting total queueing delay is then considered in addition to the cumulative delay caused by signal operations. We demonstrate the potential of this approach through microscopic traffic simulation of a real-world road network, showing a 10 − 15% reduction in average wait times over the schedule-driven traffic signal control system in heavy traffic scenarios.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
