Patent Application
20190367025
This disclosure relates to vehicle adaptive cruise control, and in particular adaptive longitudinal control using reinforcement learning.
Adaptive Cruise Control (ACC) is an important feature of autonomous driving where the vehicle is designed to cruise at a speed both efficiently and safely over a wide range of scenarios and contexts. Numerous approaches for ACC have been proposed over the years. Such approaches can be divided in two major classes: 1) an environment-centric model that primarily considers the physical signals of the environment; and 2) a driver-centric ACC model that considers the human perspective as the main design principle.
The Gazis-Herman-Rothery (GHR) model was one of the first environment-centric models proposed. This was a linear-proportional control where the desired acceleration was some factor of the velocity difference between the ego vehicle and the vehicle ahead. Many variations of this simple model were suggested to address the limitation, such as asymmetry between acceleration and deceleration, the non-zero reaction time of the driver, and interaction with multiple vehicles. In order to address the fact that each driver may have a notion of safe-distance (that might vary with the speed of an ego vehicle), a different class of models was proposed. The most well known model of this category is the Intelligent-driver model (IDM). Another approach in this class of models is where the driver reacts to the distance between the vehicles rather than the relative speed. In yet another approach, an assumption of optimal safe velocity is made. Various enhancements, including some even fairly complex enhancements, have been proposed for each of these three sub-classes. Consequently, there have also been models that prioritized simplicity, such as that identified that ACC should only consider two modes of driving: 1) in congestion (e.g., a high vehicle density area), and 2) when the ego vehicle is free of any obstacles ahead (e.g. a low vehicle density area and/or an area without other obstacles). ACC behavior may be considered as an automation where the switching happens at desired contexts.
One of the biggest flaws with the environment-centric models is the presumption that the driver is perfectly rational, time-invariant, can perceive all signals without failure, and always seeks to optimize the control. None of these is actually true in the real world. Thus, there are classes of ACC design where the driver's perspective is central, such as when a threshold to sensory perception or the size of the objects or visual angles are presumed.
Recent advancements in machine learning have fueled numerous applications in academia and various industries. In particular, supervised learning problems, such as image classification and object detection, have matured to a product-level performance. On the downside, these approaches are limited by the requirements of well-specified labelled data, which are either too expensive or too complex to obtain. Reinforcement learning, unlike supervised learning, is not limited to classification or regression problems, but can be applied to any learning problem under uncertainty and lack of knowledge of the dynamics. The approach has been applied to numerous cases where the environment model is unknown, such as with humanoids, in games, in financial markets, and many others.
Since more than a decade, high-end vehicles have been equipped with adaptive cruise control (ACC), typically through carefully designed model-based controllers. However, autonomous vehicles may require an ACC with intelligent and adaptive methods for longitudinal control of the vehicles.
This section provides a general summary of the present disclosure and is not a comprehensive disclosure of its full scope or all of its features, aspects, and objectives.
An aspect of the disclosed embodiments is a system for controlling a vehicle. The system includes a first sensor for detecting at least one environment characteristic. The system also includes a driver characteristic input device configured to receive at least one driver characteristic corresponding to a driving style of a driver. The system also includes a controller that includes a reinforcement learning adaptive cruise control that is in communication with the first sensor and the driver characteristic input device, the reinforcement learning adaptive cruise control being configured to: determine a target behavior for the vehicle based on the at least one environment characteristic and the at least one driver characteristic; and selectively control the vehicle based on the target behavior.
Another aspect of the disclosed embodiments is a longitudinal control apparatus for controlling a vehicle. The apparatus includes a processor and at least one memory. The processor is in communication with the at least one memory. The at least one memory includes instructions that, when executed by the processor, cause the processor to: receive, from a first sensor, at least one environment characteristic; receive, from a second sensor, at least one other environment characteristic; generate a homogenized environment characteristic based on the at least one environment characteristic and the at least one other environment characteristic using sensor fusion; receive, from a driver characteristic input device, at least one driver characteristic corresponding to a driving style; and determine a target behavior for the vehicle based on the homogenized environment characteristic and the at least one driver characteristic.
These and other aspects of the present disclosure are disclosed in the following detailed description of the embodiments, the appended claims, and the accompanying figures.
The disclosure is best understood from the following detailed description when read in conjunction with the accompanying drawings. It is emphasized that, according to common practice, the various features of the drawings are not to-scale. On the contrary, the dimensions of the various features are arbitrarily expanded or reduced for clarity.
The following description is merely exemplary in nature and is not intended to limit the disclosure in its application or uses. For purposes of clarity, the same reference numbers are used in the description and drawings to identify similar elements.
Based on the experimental results of using a reinforcement learning adaptive cruise control (RL-ACC) on automotive-grade vehicle simulators, a reinforcement learning based approach provides better control of autonomous vehicles than the traditional ACC. The RL-ACC may require minimal domain knowledge, is intuitive in its design, can accommodate uncertainties, can mimic human-like behavior, and may enable human-trust in the automated system. All these aspects are crucial for a fully autonomous vehicle.
In one embodiment, a system uses a general approach to ACC that is significantly different from these two classes of ACC design. The system may include a processor and a memory. The processor may be configured to execute instructions stored on the memory. For example, the memory may include instructions that, when executed by the processor, cause the processor to perform the methods described herein. The processor may be in communication with the ACC. In some embodiments, the ACC may include the processor. The processor may be configured to selectively control the ACC, as will be described. As will be described, the ACC may include a reinforcement learning adaptive cruise control.
Because the ACC design is a highly context-dependent control problem, the system incorporates changes due to both the environment and the driving styles. For example, the system may receive one or more environment characteristics from one or more sensors, one or more remotely located computing devices (e.g., cloud computing devices, mobile devices, or other suitable computing devise not directly connected to the processor), one or more weather stations, other suitable devices or systems, or a combination thereof.
Additionally, or alternatively, the system may receive one or more driver characteristics, which correspond to a driving style and/or driving preference of the driver of the vehicle. The driver characteristics may be received from one or more sensors, one or more input devices. The one or more input devices may include input devices disposed on an interior portion of the vehicle, such as a steering wheel of the vehicle, a dash of the vehicle, a display screen within the vehicle, or other input device located within the vehicle. Additionally, or alternatively, the one or more input devices may include a mobile device, or a remotely located computing device that may be configured to analyze data received from sensors or other output devices of the vehicle and determines various driving styles of the driver. For example, the driver of the vehicle may provide input using an input device to indicate a preferred driving style. Additionally, or alternatively, the one or more sensors may detect various driving style characteristics (e.g., acceleration characteristics, deceleration characteristics, turning and/or steering control characteristics, other suitable characteristics, or a combination thereof) and provide the driving style characteristics to the processor. Additionally, or alternatively, the driving characteristic may correspond to a driving style of a driver of another vehicle. For example, the processor may receive the driving characteristics from a remotely located computer (e.g., via a vehicle-to-vehicle or vehicle-to-X, communication). The processor may determine a driving style of the driver of another vehicle using the driving characteristic. Additionally, or alternatively, the driving characteristic may correspond to a driving style of an autonomous vehicle. For example, the vehicle and/or the other vehicle may include autonomous vehicles. The processor may receive the driving characteristic from sensors of the vehicle and/or from a remotely located computer (e.g., via a vehicle-to-vehicle or vehicle-to-X, communication) corresponding to another vehicle. The driving characteristic may correspond to a driving style of the vehicle and/or the other vehicle. As will be described, the processor is configured to determine a target behavior (e.g., which may be referred to as an optimal behavior or a behavior) for the vehicle and to selectively control the vehicle based on the target behavior.
Instead of aiming for an optimal control for ACC, the system can create a human-like cruise driving that can adapt over time. Incorporating such sources of uncertainty increases the model complexity. To address such complexity, the system can use model-free approaches of artificial intelligence where exact modelling of the environment is skirted through efficient use of data. Additionally, or alternatively, the system can allow for truly adaptive cruise control that adapts not just to the environment but also the changing preferences of the driver over time.
The longitudinal control of an autonomous vehicle has two major conflicting goals: 1) to maintain as close to the set speed as possible, and 2) to have as much safe distance from the preceding vehicle as possible. The longitudinal control may also include additional objectives, such as maintaining an optimally “smooth” trajectory. Because ACC can be used as a comfort feature, in actual deployment of ACC in the vehicle, additional objectives are very crucial. In a real-world scenario, longitudinal control of the vehicle is seldom, if ever, decoupled from other aspects such as lateral control, traffic situations, and behavior of the other vehicles. However, these factors, when considered together, make the problem complex.
The systems and methods used to overcome such a complex problem are described herein. The ego vehicle is denoted by E and the vehicle ahead of the ego vehicle is denoted by A. The position of the ego vehicle is denoted by x and the velocity and acceleration are denoted by {dot over (x)} and {umlaut over (x)}, respectively. Furthermore, separation between these vehicles is denoted by Δx, i.e., Δx:=xA−xε. The ego vehicle can be controlled by an action, such as a control action, which is denoted by uε. For example, the control action can be a new position of the vehicle. Acceleration (e.g., using a gas pedal) and deceleration (e.g., using a brake) is used for longitudinal control. The control action therefore can be acceleration, i.e., uεε[−bmax, amax]Q, where bmax is the maximum deceleration and amax is the maximum acceleration that could be achieved given the physics of the vehicle. Because the precision of actual control of acceleration (or deceleration) can be governed by the dynamics of the vehicle, the control action range can be defined over rational numbers, i.e., uεεQ. Comfortable acceleration and deceleration limits are usually a strict subset of this range. Therefore, the function ƒθ(Δx): R→R denotes the objective of maintaining a safe distance. Usually such a function would be parametrized by requirements, such as a safe distance parameter. These parameters are denoted by θ. Similarly, the function gθ({dot over (x)}ε): R→R stands for the objective of maintaining the set speed. Because a given situation or driver preference determines such a set speed, this is subsumed in the parameter θ. The dynamics of the environment is a black-box function D(t), which at any instant t>0 provides all necessary kinematics values, i.e., xε, {dot over (x)}ε, {circumflex over (x)}ε, xA, {dot over (x)}A and {umlaut over (x)}A; which are denoted by X.
Given the positions of the vehicles and possibly their time derivatives, classical-ACC can be defined as a control problem of finding the optimal acceleration action that maintains a desired safe distance while maintaining a set speed. Mathematically, classical-ACC is equivalent to the following functions:
Minimize uεƒθ(Δx)+λgθ({dot over (x)}E)
subject to uεϵ[—bmax,amax]
∀t>0,Xt=D(t)
Here, λ∈R can facilitate the convex combination of the conflicting objectives of maintaining a set-speed and a safe distance. In general, such a combination may also be some nonlinear function. In this example, a linear combination is used for simplicity. The classical-ACC does not consider the driver preferences explicitly. In fact, a crucial limitation of such models is that the parameters do not have intuitive meaning when actual traffic data is considered. In other words, the parameters of the classical-ACC models cannot be set through observing actual traffic data and human behavior.
Several approaches are discussed in more detail herein. The first approach is a generalized approach to longitudinal control. Similar to the approaches mentioned above, the longitudinal control creates a problem separate from other factors such as lateral control, traffic conditions, and behaviors of the other drivers. In one embodiment, the method of the system is a function of only longitudinal variables.
The control problem involves controlling the ego vehicle in the longitudinal direction when the position of the vehicle ahead is known through some sensory input. Unlike the model-based approach, the physical model of the problem and its position may be provided in nonmetric terms, such as through a viewing angle, the image itself, or a laser scan. Radar measurements may be used as the inputs. The input variables are represented by the vector Z. In a simple scenario, this would be all kinematics variables provided through ideal sensors, i.e., Z=X. The control strategy can be used to map a given observation to the desirable action under certain behavior, i.e., π(Z)=uε.
The intuitive behavior model of the controller can be similar to those vehicle following models that consider human aspects of driving. However, unlike those approaches, the system models all the human-like preferences and biases them under a single framework of a reward schema. For example, a conservative driver may put a higher negative reward for an unsafe close distance than for a usual driver. Similarly, in a scenario of a driver on a highway of high-speed vehicles, faster speeds may carry a greater reward.
Longitudinal (and lateral) control of the vehicle by human drivers can typically be in the order of few seconds. For example, once a driver commits to accelerating towards the vehicle ahead, the driver is not likely to change the acceleration within the next few milliseconds. Considering the abstract-level of human-like control of the vehicle greatly signifies the learning problem. This is similar to hybrid control strategy where the high-level controller is responsible for discrete jumps between low-level continuous state controllers. Similar to such approaches, a low-level controller can be used. In the industrial setting, such controllers can be designed based on specifics of the vehicle and their numerous physical parameters.
In order to obtain a generalized longitudinal controller, an artificial intelligence approach can be used with the controller design. In particular, reinforcement learning can be used to learn the controller through various active simulations. On one hand, this allows the system to harness the reward schema. On the other hand, it is a general approach that degenerates into known model-based controllers under suitable conditions.
The learning approach can be used to generalize the classical-ACC problem into one where control is learned through the interaction with the world. Given the observations of the world Z, the learned adaptive cruise control RL ACC can be defined as a learning problem of finding the near-optimal policy π (of acceleration control) under a given set of rewards specified for different scenes and actions. For example, Rπ(Z0), where Z0∈Z is the initial state at time t=0, where after the policy, π is followed and Z is set of all possible initial states. Mathematically, RL-ACC is equivalent to following functions:
This approach does not require explicitly setting the functions ƒθ(Δx) and gθ({dot over (x)}ε). In situations where these functions are known, the system can incorporate them into the rewarding schema.
In one embodiment, the system uses a least square policy iteration. Markovian Decision Problem (MDP) involves a probabilistic system represented by the tuple (S, U, T, R, γ). Here, S is the countable set of states of the system and U is the action space. The uncertain nature of the environment is encoded in the transition system T: S×S→R[0,1]. The rewards in each state (and action) can be determined by R: S→R, and γ∈R[0,1] is the discounting factor. MDP is an NP-hard problem of finding an optimal policy π*: S→U, which maximizes the total discounted reward. One of the approaches to achieve this approximately is to define Q-value as:
By applying the Bellman optimality criteria—i.e., the Q-value at any step t with the state s and the action u returned by an optimal policy is the same as the optimal reward obtained at this step and following the optimal policy for all subsequent steps (t+1 onward)—the system can compute the Q-value under the current policy. This can be identified as a policy evaluation. In order to improve a policy, the system may select the actions that maximize the computed Q-value. This step is policy improvement. When performed iteratively over the space of deterministic policies, policy evaluation and policy improvement result in optimal policy. This is the essence of policy iteration. However, under this algorithm, convergence is guaranteed only when the representation of the Q-values, as well as the policy, are exact such as with tabular form. For large (or infinite) state space, both these representations need to be parametrized and hence may be only approximate.
The system can use a least square Policy Iteration (LSPI). LSPI is a class of approximate policy iteration reinforcement learning algorithms. The actual policy is projected to an approximate space such that the iterative operator yields Q-values that are near optimal with respect to an actual fixed point in the sense of L2-norm.
The system can use an approach for learning the ACC through LSPI. The system can first generate the data of the interaction of the vehicle with the environment. This can be performed through third-party simulators. Once the required data of interaction is available, the system applies the RL algorithm as described in the first algorithm described below. The system devises the reward schema based on the control and the environment. The system may use another domain specific aspect, such as the choice of Gaussians. The system can use the Gaussians to approximate the actual Q-value. The centers of these Gaussians can be used at the interesting parts of the state-space. Based on the number of these Gaussians, the system can construct a random weight vector. Once the system accomplishes the problem design, the system uses the RL-ACC to obtain optimum weight vector. The system can use the following Algorithm 1 for RL-ACC:
The system can use a cruise control mode. A cruise control functionality can be designed for a specific model in order to maintain a constant velocity when the ego vehicle can drive freely. This feature is primarily based on a proportional-integral-derivative (PID) controller. Through multiple simulations, the system can set the values of the proportional tuning constant (KP) and the derivative constant (Kd), for example, to 10.0 each. A value of 10.0 is stable for the optimal smoothness at different velocities of the ego vehicle. The output of the PID controller may be limited to the maximum and minimum acceleration values. The desired acceleration can be calculated through the following equation:
where subscripts cruise and curr stand for kinematics of the ego vehicle for cruising and current state respectively. The system can obtain the acceleration control via the PID controller. Δcte stands for cross track error and the time-step is dt=40 ms.
The system can combine the ACC and cruise control modes. The system can switch between the cruise mode and the ACC mode based on the threshold ACC distance (TAD), the ACC mode, and the safety mode based on the threshold safety distance (TSD) between the ego vehicle and the lead vehicle set by a user. The inter vehicle distance (IVD) can be a function of the long-range radar output from the ego vehicle.
Extensive experiments were carried out on the system to evaluate the RL-ACC empirically. For example,
The system can use different simulators, each of which are important for different reasons. Two simulators, besides other mathematical software tests and closed loop tests, are described in more detail herein. As illustrated in
As described below, by way of example, various software and hardware pipelines can be utilized with the systems and methods. As for software, the learning algorithm can be implemented in MATLAB or other suitable software for ease of prototyping whereas during the actual deployment the learned policy can be in native C++ code or other suitable code on a distributed system set-up on a Robot Operating System (ROS). The software stack can be composed of several ROS nodes intercommunicating with each other at different levels based on their individual functionalities. The ROS nodes can communicate to the simulators through a Transmission Control Protocol/Internet Protocol (TCP/IP) network configuration, or other suitable configuration. The system can include additional and/or fewer software and hardware components and is not limited to those described herein.
The system can include various hardware. For example, the entire software stack can run on the DriveCore™ platform, as illustrated in
The system can handle both cases of ACC applications—on highway with high set-speed and in urban environment with low set-speed. Even in the simple set-up of one leading vehicle as illustrated in
The system can use the RL-ACC in different scenarios. Like any other model of ACC, a criterion of success is for the ego vehicle to follow the lead vehicle at a safe distance while maintaining a smooth velocity profile. Besides that, the system also compares velocity profiles with empirical models of ACC like the IDM. Finally, the dynamic nature of RL-ACC during aggressive and conservative behaviors of the controller is compared in the same scenario.
By way of example, for single-vehicle scenarios, experiments on the system can be conducted in Oktal. Such single-vehicle scenarios may involve simple scenarios with relatively ideal conditions. In a first experiment, three lanes (versus the two lanes shown in
In this experiment, the lead vehicle maintains a constant velocity throughout its journey. The main objective here is to test the very basic expectation of any ACC model, i.e., for the ego vehicle to be able to adjust its velocity based on the preceding vehicle in the simplest possible scenario with no other agents involved. The velocity profile illustrated in
In another exemplary scenario, the lead vehicle having a variable velocity is tested. A vehicle using the RL-ACC system can follow vehicles at a constant velocity while maintaining a safe distance. Hence, the second experiment includes the scenario with a higher level of complexity. In this second experiment, the lead vehicle changes its velocity after crossing a preset checkpoint on the simulator track, resulting in the ACC model to adapt to multiple velocities over the course of the journey.
In another exemplary scenario, the velocity profile in single-vehicle scenarios is evaluated. A test of uncertainty, which is a sharp rise/drop in velocity of the lead vehicle in short ranges, can be used in such scenarios. For example, vehicles can break away from their steady motion and either accelerate/decelerate abruptly in both highways and country roads. The system can adapt to these situations. Obviously, the system can jump back and forth between the cruise and the ACC modes frequently. In this example, the lead vehicle changes its velocity to 13.89 m/s after crossing checkpoint A, moving away from the ego vehicle rapidly. As the inter-vehicle distance becomes greater than 140 m, the ego vehicle switches back to the cruise mode like initially and starts accelerating. Post checkpoint B, as the lead vehicle slows down, the ACC starts decelerating. The experiment evaluates whether the system can handle a sharp drop in velocity (5.55 m/s) over a very short distance while maintaining a safe distance with the lead vehicle.
In another exemplary scenario, the multi-vehicle scenarios is evaluated. Vires can be used for conducting experiments in more complex scenarios like realistic highway and residential roads with congested traffic while comparing with empirical models of ACC like the IDM. Two different scenarios were designed for these experiments: 1) crowded highway on the German Autobahn with a relatively higher congestion of vehicles (
While it is relatively easier for human drivers to react to lane changes/merges with other vehicles cutting in/out (e.g. on the crowded highway with lane changes), it might be more challenging for the ACC to adapt in real-time, being dependent on the sensor fusion of outputs of both radars. Instead of reacting to this output directly, the system considers it over consecutive time steps, which enables smoother control and makes the learner context-aware. Hence, when the lead vehicle changes or is absent in the new lane, the ACC can reset to either the cruise mode (lead vehicle absent) or adjust its velocity based on the new lead vehicle's velocity in a short span of time. In this example, the system is tested on a long stretch of the Autobahn while manually changing lanes a number of times.
In another exemplary scenario, conservative versus aggressive RL-ACC on countryside roads is evaluated. An advantage of the RL-ACC over IDM is RL-ACC's ability to adapt to different environments and driving conditions. Ideally, ACC should behave differently in highway scenarios and city roads. In the former, it should behave more conservatively, i.e., the distance between vehicles should always be relatively large because of the average speed, whereas this distance would be a lot less in the later, resulting in a more aggressive behavior of the ACC. In this experiment, the system behaves in a conservative and an aggressive manner on countryside roads.
Typically, ACC is approached as a model-based controller design. The systems and methods described herein fix the control problems of ACC by using artificial intelligence techniques. In particular, the systems and methods use RL-ACC. While the classical approach—IDM—is widely used as the go-to ACC model in the automotive industry and performs decently on the highways, due to its model-based approach, it lacks the ability to adapt to the environments or driving preferences. Because the RL-ACC does not require domain knowledge, it can be trained irrespective of the environment. Moreover, the systems and methods using this approach also make the controller design intuitive from the perspective of human driving. Complex controls can arise out of rather simple rewarding strategies. This enhances both the reasoning of the changes in control of the vehicle as well as trust of human passengers when the autonomous vehicle is utilizing ACC. The systems and methods can create human-like autonomous driving.
Furthermore, the systems and methods provide the safety of learned control and devise a general human-like controller for autonomous driving. In order to achieve the former, systematic verification of the controller can be performed using state-of-the-art formal methods. For achieving the latter goal, one of the approaches is to invoke the deep learning methods that can ingest the input images and learn ACC as some sort of optimal behavior in different contexts.
In some embodiments, a system for controlling a vehicle includes a first sensor that detects at least one environment characteristic. The system also includes a driver characteristic input device configured to receive at least one driver characteristic corresponding to a driving style of a driver of the vehicle. The system also includes a reinforcement learning adaptive cruise control configured to: determine a target behavior for the vehicle based on the at least one environment characteristic and the at least one driver characteristic; and selectively control the vehicle based on the target behavior.
In some embodiments, the at least one environment characteristic includes a weather condition proximate the vehicle. In some embodiments, the at least one environment characteristic indicates that the vehicle is being driven in a high vehicle density area. In some embodiments, the at least one environment characteristic indicates that the vehicle is being driven in a low vehicle density area. In some embodiments, the driver characteristic input device is disposed on an interior portion of the vehicle. In some embodiments, the driver characteristic input device is disposed on mobile computing device. In some embodiments, the at least one driver characteristic indicates that the driver is a conservative driver. In some embodiments, the at least one driver characteristic indicates that the driver is an aggressive driver. In some embodiments, the first sensor includes a short-range radar. In some embodiments, the system also includes a second sensor configured to detect at least one other environment characteristic. In some embodiments, the second sensor includes a long-range radar, and wherein the reinforcement learning adaptive cruise control is further configured to use sensor fusion to generate a homogenized environment characteristic based on the at least one environment characteristic and the at least one other environment characteristic. In some embodiments, the reinforcement learning adaptive cruise control is further configured to determine the target behavior for the vehicle based on the homogenized environment characteristic and the at least one driver characteristic.
In some embodiments, a longitudinal control apparatus for controlling a vehicle includes a processor and at least one memory. The processor is in communication with the at least one memory. The at least one memory includes instructions that, when executed by the processor, cause the processor to: receive, from a first sensor, at least one environment characteristic; receive, from a second sensor, at least one other environment characteristic; generate a homogenized environment characteristic based on the at least one environment characteristic and the at least one other environment characteristic using sensor fusion; receive, from a driver characteristic input device, at least one driver characteristic corresponding to a driving style of a driver of the vehicle; determine a target behavior for the vehicle based on the homogenized environment characteristic and the at least one driver characteristic; and selectively control the vehicle based on the target behavior.
In some embodiments, the first sensor includes a short-range radar and the second sensor includes a long-rage radar. In some embodiments, the at least one environment characteristic includes a weather condition proximate the vehicle. In some embodiments, the at least one environment characteristic indicates that the vehicle is being driven in a high vehicle density area. In some embodiments, the at least one environment characteristic indicates that the vehicle is being driven in a low vehicle density area. In some embodiments, the driver characteristic input device is disposed on an interior portion of the vehicle. In some embodiments, the driver characteristic input device is disposed on mobile computing device. In some embodiments, the at least one driver characteristic indicates that the driver is a conservative driver.
In some embodiments, a system for controlling a vehicle includes a first sensor for detecting at least one environment characteristic. The system also includes a driver characteristic input device configured to receive at least one driver characteristic corresponding to a driving style of a driver. The system also includes a controller that includes a reinforcement learning adaptive cruise control that is in communication with the first sensor and the driver characteristic input device, the reinforcement learning adaptive cruise control being configured to: determine a target behavior for the vehicle based on the at least one environment characteristic and the at least one driver characteristic; and selectively control the vehicle based on the target behavior.
In some embodiments, the at least one environment characteristic includes a weather condition proximate the vehicle. In some embodiments, the at least one environment characteristic indicates that the vehicle is being driven in a high vehicle density area. In some embodiments, the at least one environment characteristic indicates that the vehicle is being driven in a low vehicle density area. In some embodiments, the driver characteristic input device is disposed on an interior portion of the vehicle. In some embodiments, the driver characteristic input device is disposed on a mobile computing device. In some embodiments, the at least one driver characteristic indicates that the driver is a conservative driver. In some embodiments, the at least one driver characteristic indicates that the driver is an aggressive driver. In some embodiments, the first sensor includes a short-range detection device. In some embodiments, the system also includes a second sensor configured to detect at least one other environment characteristic. In some embodiments, the second sensor includes a long-range detection device, and wherein the reinforcement learning adaptive cruise control is further configured to use sensor fusion to generate a homogenized environment characteristic based on the at least one environment characteristic and the at least one other environment characteristic. In some embodiments the reinforcement learning adaptive cruise control is further configured to determine the target behavior for the vehicle based on the homogenized environment characteristic and the at least one driver characteristic.
In some embodiments, a longitudinal control apparatus for controlling a vehicle includes a processor and at least one memory. The processor is in communication with the at least one memory. The at least one memory includes instructions that, when executed by the processor, cause the processor to: receive, from a first sensor, at least one environment characteristic; receive, from a second sensor, at least one other environment characteristic; generate a homogenized environment characteristic based on the at least one environment characteristic and the at least one other environment characteristic using sensor fusion; receive, from a driver characteristic input device, at least one driver characteristic corresponding to a driving style; and determine a target behavior for the vehicle based on the homogenized environment characteristic and the at least one driver characteristic.
In some embodiments, the first sensor includes a short-range detection device and the second sensor includes a long-rage detection device. In some embodiments, the driving style corresponds to a driving style of a driver. In some embodiments, the driving style corresponds a driving style of an autonomous vehicle. In some embodiments, the processor is further configured to mimic human control of the vehicle using the target behavior for the vehicle. In some embodiments, the processor is further configured to control the vehicle to follow at least one other vehicle at a constant velocity while maintaining a safe distance between the vehicle and the at least one other vehicle. In some embodiments, the processor is further configured to conservatively control the vehicle in response to the driving style indicating a conservative driving style. In some embodiments, the processor is further configured to aggressively control the vehicle in response to the driving style indicating an aggressive driving style.
While the disclosure has been described in connection with certain embodiments, it is to be understood that the disclosure is not to be limited to the disclosed embodiments but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the scope of the appended claims, which scope is to be accorded the broadest interpretation so as to encompass all such modifications and equivalent structures as is permitted under the law.
The above discussion is meant to be illustrative of the principles and various embodiments of the present invention. Numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated. It is intended that the following claims be interpreted to embrace all such variations and modifications.
Various terms are used to refer to particular system components. In the above discussion and in the claims, the terms “including” and “comprising” are used in an open-ended fashion, and thus should be interpreted to mean “including, but not limited to . . . .” Also, the term “couple” or “couples” is intended to mean either an indirect or direct connection. Thus, if a first device couples to a second device, that connection may be through a direct connection or through an indirect connection via other devices and connections.
“Controller” shall mean individual circuit components, an application-specific integrated circuit (ASIC), a microcontroller with controlling software, a digital signal processor (DSP), a processor with controlling software, a field programmable gate array (FPGA), or combinations thereof.
The word “example” is used herein to mean serving as an example, instance, or illustration. Any aspect or design described herein as “example” is not necessarily to be construed as preferred or advantageous over other aspects or designs. Rather, use of the word “example” is intended to present concepts in a concrete fashion. As used in this application, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or.” That is, unless specified otherwise, or clear from context, “X includes A or B” is intended to mean any of the natural inclusive permutations. That is, if X includes A; X includes B; or X includes both A and B, then “X includes A or B” is satisfied under any of the foregoing instances. In addition, the articles “a” and “an” as used in this application and the appended claims should generally be construed to mean “one or more” unless specified otherwise or clear from context to be directed to a singular form. Moreover, use of the term “an implementation” or “one implementation” throughout is not intended to mean the same embodiment or implementation unless described as such.
Implementations the systems, algorithms, methods, instructions, etc., described herein can be realized in hardware, software, or any combination thereof. The hardware can include, for example, computers, intellectual property (IP) cores, application-specific integrated circuits (ASICs), programmable logic arrays, optical processors, programmable logic controllers, microcode, microcontrollers, servers, microprocessors, digital signal processors, or any other suitable circuit. In the claims, the term “processor” should be understood as encompassing any of the foregoing hardware, either singly or in combination. The terms “signal” and “data” are used interchangeably.
As used herein, the term module can include a packaged functional hardware unit designed for use with other components, a set of instructions executable by a controller (e.g., a processor executing software or firmware), processing circuitry configured to perform a particular function, and a self-contained hardware or software component that interfaces with a larger system. For example, a module can include an application specific integrated circuit (ASIC), a Field Programmable Gate Array (FPGA), a circuit, digital logic circuit, an analog circuit, a combination of discrete circuits, gates, and other types of hardware or combination thereof. In other embodiments, a module can include memory that stores instructions executable by a controller to implement a feature of the module.
Further, in one aspect, for example, systems described herein can be implemented using a general-purpose computer or general-purpose processor with a computer program that, when executed, carries out any of the respective methods, algorithms, and/or instructions described herein. In addition, or alternatively, for example, a special purpose computer/processor can be utilized which can contain other hardware for carrying out any of the methods, algorithms, or instructions described herein.
Further, all or a portion of implementations of the present disclosure can take the form of a computer program product accessible from, for example, a computer-usable or computer-readable medium. A computer-usable or computer-readable medium can be any device that can, for example, tangibly contain, store, communicate, or transport the program for use by or in connection with any processor. The medium can be, for example, an electronic, magnetic, optical, electromagnetic, or a semiconductor device. Other suitable mediums are also available.
The above-described embodiments, implementations, and aspects have been described in order to allow easy understanding of the present invention and do not limit the present invention. On the contrary, the invention is intended to cover various modifications and equivalent arrangements included within the scope of the appended claims, which scope is to be accorded the broadest interpretation to encompass all such modifications and equivalent structure as is permitted under the law.
This application claims the benefit of and priority to U.S. Provisional Application No. 62/678,297, filed May 31, 2018, which is incorporated herein by reference in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| 62678297 | May 2018 | US |
