Sustained sensorimotor control as intermittent decisions about prediction errors: computational framework and application to ground vehicle steering

A model of dyadic merging interactions explains human drivers' behavior from control inputs to decisions

Safe and socially acceptable interactions with human-driven vehicles are a major challenge in automated driving. A good understanding of the underlying principles of such traffic interactions could help address this challenge. Particularly, accurate driver models could be used to inform automated vehicles in interactions. These interactions entail complex dynamic joint behaviors composed of individual driver contributions in terms of high-level decisions, safety margins, and low-level control inputs. Existing driver models typically focus on one of these aspects, limiting our understanding of the underlying principles of traffic interactions. Here, we present a Communication-Enabled Interaction model based on risk perception, that does not assume humans are rational and explicitly accounts for communication between drivers. Our model can explain and reproduce observed human interactions in a simplified merging scenario on all three levels. Thereby improving our understanding of the underlying mechanisms of human traffic interactions and posing a step towards interaction-aware automated driving.

Should I Stay or Should I Go? Cognitive Modeling of Left-Turn Gap Acceptance Decisions in Human Drivers

OBJECTIVE

We aim to bridge the gap between naturalistic studies of driver behavior and modern cognitive and neuroscientific accounts of decision making by modeling the cognitive processes underlying left-turn gap acceptance by human drivers.

BACKGROUND

Understanding decisions of human drivers is essential for the development of safe and efficient transportation systems. Current models of decision making in drivers provide little insight into the underlying cognitive processes. On the other hand, laboratory studies of abstract, highly controlled tasks point towards noisy evidence accumulation as a key mechanism governing decision making. However, it is unclear whether the cognitive processes implicated in these tasks are as paramount to decisions that are ingrained in more complex behaviors, such as driving.

RESULTS

The drivers' probability of accepting the available gap increased with the size of the gap; importantly, response time increased with time gap but not distance gap. The generalized drift-diffusion model explained the observed decision outcomes and response time distributions, as well as substantial individual differences in those. Through cross-validation, we demonstrate that the model not only explains the data, but also generalizes to out-of-sample conditions.

CONCLUSION

Our results suggest that dynamic evidence accumulation is an essential mechanism underlying left-turn gap acceptance decisions in human drivers, and exemplify how simple cognitive process models can help to understand human behavior in complex real-world tasks.

APPLICATION

Potential applications of our results include real-time prediction of human behavior by automated vehicles and simulating realistic human-like behaviors in virtual environments for automated vehicles.

Brownian processes in human motor control support descending neural velocity commands

The motor neuroscience literature suggests that the central nervous system may encode some motor commands in terms of velocity. In this work, we tackle the question: what consequences would velocity commands produce at the behavioral level? Considering the ubiquitous presence of noise in the neuromusculoskeletal system, we predict that velocity commands affected by stationary noise would produce "random walks", also known as Brownian processes, in position. Brownian motions are distinctively characterized by a linearly growing variance and a power spectral density that declines in inverse proportion to frequency. This work first shows that these Brownian processes are indeed observed in unbounded motion tasks e.g., rotating a crank. We further predict that such growing variance would still be present, but bounded, in tasks requiring a constant posture e.g., maintaining a static hand position or quietly standing. This hypothesis was also confirmed by experimental observations. A series of descriptive models are investigated to justify the observed behavior. Interestingly, one of the models capable of accounting for all the experimental results must feature forward-path velocity commands corrupted by stationary noise. The results of this work provide behavioral support for the hypothesis that humans plan the motion components of their actions in terms of velocity.

Dynamic primitives in constrained action: systematic changes in the zero-force trajectory

Humans substantially outperform robotic systems in tasks that require physical interaction, despite seemingly inferior muscle bandwidth and slow neural information transmission. The control strategies that enable this performance remain poorly understood. To bridge that gap, this study examined kinematically constrained motion as an intermediate step between the widely studied unconstrained motions and sparsely studied physical interactions. Subjects turned a horizontal planar crank in two directions (clockwise and counterclockwise) at three constant target speeds (fast, medium, and very slow) as instructed via visual display. With the hand constrained to move in a circle, nonzero forces against the constraint were measured. This experiment exposed two observations that could not result from mechanics alone but may be attributed to neural control composed of dynamic primitives. A plausible mathematical model of interactive dynamics (mechanical impedance) was assumed and used to "subtract" peripheral neuromechanics. This method revealed a summary of the underlying neural control in terms of motion, a zero-force trajectory. The estimated zero-force trajectories were approximately elliptical and their orientation differed significantly with turning direction; that is consistent with control using oscillations to generate an elliptical zero-force trajectory. However, for periods longer than 2-5 s, motion can no longer be perceived or executed as periodic. Instead, it decomposes into a sequence of submovements, manifesting as increased variability. These quantifiable performance limitations support the hypothesis that humans simplify this constrained-motion task by exploiting at least three primitive dynamic actions: oscillations, submovements, and mechanical impedance.NEW & NOTEWORTHY Control using primitive dynamic actions may explain why human performance is superior to robots despite seemingly inferior "wetware"; however, this also implies limitations. For a crank-turning task, this work quantified two such informative limitations. Force was exerted even though it produced no mechanical work, the underlying zero-force trajectory was roughly elliptical, and its orientation differed with turning direction, evidence of oscillatory control. At slow speeds, speed variability increased substantially, indicating intermittent control via submovements.

Applying the Accumulator model to predict driver's reaction time based on looming in approaching and braking conditions

INTRODUCTION

The prediction of when the driver will react to a change in the lead vehicle motion is critical for assessing rear-end crash risk using car-following models. Past studies have assumed constant reaction time and driver's continuous reaction. However, these assumptions are not valid as the driver's reaction time can vary in different car-following situations and the driver does not continuously react to the lead vehicle motion. Thus, this study predicted the driver's reaction time using the Wiedemann car-following model and the Accumulator model. The Accumulator model assumes the driver's start of reaction based on the accumulation of looming and thereby reflects the driver's intermittent reaction.

METHOD

Fifty drivers' behavior was observed using a driving simulator in two scenarios: (1) approach and follow a moving lead vehicle and (2) approach a stopped lead vehicle. The Accumulator model predicted the reaction times based on different looming variables (angular velocity and tau-inverse), lead vehicle type (car and truck), and lead vehicle brake lights (on or off).

RESULTS

The Accumulator model showed lower prediction errors of the reaction time than the Wiedemann model, which assumes reaction based on the fixed looming threshold. The Accumulator model predicted the reaction times more accurately when it was calibrated with the angular velocity due to width and height of lead vehicles. Moreover, the Accumulator model with tau-inverse produced the smallest prediction error of reaction times among different Accumulator models and the Wiedemann model when lead vehicle brake lights were on.

CONCLUSIONS

This study demonstrates that the Accumulator model is a promising method of predicting the driver's reaction time in car-following situations, which affects rear-end crash risk.

PRACTICAL APPLICATIONS

The Accumulator model can be incorporated into a car-following model for the prediction of reaction times and can estimate the rear-end collision risk of vehicles more accurately.

Explaining human interactions on the road by large-scale integration of computational psychological theory

When humans share space in road traffic, as drivers or as vulnerable road users, they draw on their full range of communicative and interactive capabilities. Much remains unknown about these behaviors, but they need to be captured in models if automated vehicles are to coexist successfully with human road users. Empirical studies of human road user behavior implicate a large number of underlying cognitive mechanisms, which taken together are well beyond the scope of existing computational models. Here, we note that for all of these putative mechanisms, computational theories exist in different subdisciplines of psychology, for more constrained tasks. We demonstrate how these separate theories can be generalized from abstract laboratory paradigms and integrated into a computational framework for modeling human road user interaction, combining Bayesian perception, a theory of mind regarding others' intentions, behavioral game theory, long-term valuation of action alternatives, and evidence accumulation decision-making. We show that a model with these assumptions-but not simpler versions of the same model-can account for a number of previously unexplained phenomena in naturalistic driver-pedestrian road-crossing interactions, and successfully predicts interaction outcomes in an unseen data set. Our modeling results contribute to demonstrating the real-world value of the theories from which we draw, and address calls in psychology for cumulative theory-building, presenting human road use as a suitable setting for work of this nature. Our findings also underscore the formidable complexity of human interaction in road traffic, with strong implications for the requirements to set on development and testing of vehicle automation.

A Cognitive Model to Anticipate Variations of Situation Awareness and Attention for the Takeover in Highly Automated Driving

The development of highly automated driving requires dynamic approaches that anticipate the cognitive state of the driver. In this paper, a cognitive model is developed that simulates a spectrum of cognitive processing and the development of situation awareness and attention guidance in different takeover situations. In order to adapt cognitive assistance systems according to individuals in different situations, it is necessary to understand and simulate dynamic processes that are performed during a takeover. To validly represent cognitive processing in a dynamic environment, the model covers different strategies of cognitive and visual processes during the takeover. To simulate the visual processing in detail, a new module for the visual attention within different traffic environments is used. The model starts with a non-driving-related task, attends the takeover request, makes an action decision and executes the corresponding action. It is evaluated against empirical data in six different driving scenarios, including three maneuvers. The interaction with different dynamic traffic scenarios that vary in their complexity is additionally represented within the model. Predictions show variances in reaction times. Furthermore, a spectrum of driving behavior in certain situations is represented and how situation awareness is gained during the takeover process. Based on such a cognitive model, an automated system could classify the driver’s takeover readiness, derive the expected takeover quality and adapt the cognitive assistance for takeovers accordingly to increase safety.

Computational modeling of driver pre-crash brake response, with and without off-road glances: Parameterization using real-world crashes and near-crashes

When faced with an imminent collision threat, human vehicle drivers respond with braking in a manner which is stereotypical, yet modulated in complex ways by many factors, including the specific traffic situation and past driver eye movements. A computational model capturing these phenomena would have high applied value, for example in virtual vehicle safety testing methods, but existing models are either simplistic or not sufficiently validated. This paper extends an existing quantitative driver model for initiation and modulation of pre-crash brake response, to handle off-road glance behavior. The resulting models are fitted to time-series data from real-world naturalistic rear-end crashes and near-crashes. A stringent parameterization and model selection procedure is presented, based on particle swarm optimization and maximum likelihood estimation. A major contribution of this paper is the resulting first-ever fit of a computational model of human braking to real near-crash and crash behavior data. The model selection results also permit novel conclusions regarding behavior and accident causation: Firstly, the results indicate that drivers have partial visual looming perception during off-road glances; that is, evidence for braking is collected, albeit at a slower pace, while the driver is looking away from the forward roadway. Secondly, the results suggest that an important causation factor in crashes without off-road glances may be a reduced responsiveness to visual looming, possibly associated with cognitive driver state (e.g., drowsiness or erroneous driver expectations). It is also demonstrated that a model parameterized on less-critical data, such as near-crashes, may also accurately reproduce driver behavior in highly critical situations, such as crashes.

Accumulation of continuously time-varying sensory evidence constrains neural and behavioral responses in human collision threat detection

Evidence accumulation models provide a dominant account of human decision-making, and have been particularly successful at explaining behavioral and neural data in laboratory paradigms using abstract, stationary stimuli. It has been proposed, but with limited in-depth investigation so far, that similar decision-making mechanisms are involved in tasks of a more embodied nature, such as movement and locomotion, by directly accumulating externally measurable sensory quantities of which the precise, typically continuously time-varying, magnitudes are important for successful behavior. Here, we leverage collision threat detection as a task which is ecologically relevant in this sense, but which can also be rigorously observed and modelled in a laboratory setting. Conventionally, it is assumed that humans are limited in this task by a perceptual threshold on the optical expansion rate-the visual looming-of the obstacle. Using concurrent recordings of EEG and behavioral responses, we disprove this conventional assumption, and instead provide strong evidence that humans detect collision threats by accumulating the continuously time-varying visual looming signal. Generalizing existing accumulator model assumptions from stationary to time-varying sensory evidence, we show that our model accounts for previously unexplained empirical observations and full distributions of detection response. We replicate a pre-response centroparietal positivity (CPP) in scalp potentials, which has previously been found to correlate with accumulated decision evidence. In contrast with these existing findings, we show that our model is capable of predicting the onset of the CPP signature rather than its buildup, suggesting that neural evidence accumulation is implemented differently, possibly in distinct brain regions, in collision detection compared to previously studied paradigms.

Drivers use active gaze to monitor waypoints during automated driving

Automated vehicles (AVs) will change the role of the driver, from actively controlling the vehicle to primarily monitoring it. Removing the driver from the control loop could fundamentally change the way that drivers sample visual information from the scene, and in particular, alter the gaze patterns generated when under AV control. To better understand how automation affects gaze patterns this experiment used tightly controlled experimental conditions with a series of transitions from 'Manual' control to 'Automated' vehicle control. Automated trials were produced using either a 'Replay' of the driver's own steering trajectories or standard 'Stock' trials that were identical for all participants. Gaze patterns produced during Manual and Automated conditions were recorded and compared. Overall the gaze patterns across conditions were very similar, but detailed analysis shows that drivers looked slightly further ahead (increased gaze time headway) during Automation with only small differences between Stock and Replay trials. A novel mixture modelling method decomposed gaze patterns into two distinct categories and revealed that the gaze time headway increased during Automation. Further analyses revealed that while there was a general shift to look further ahead (and fixate the bend entry earlier) when under automated vehicle control, similar waypoint-tracking gaze patterns were produced during Manual driving and Automation. The consistency of gaze patterns across driving modes suggests that active-gaze models (developed for manual driving) might be useful for monitoring driver engagement during Automated driving, with deviations in gaze behaviour from what would be expected during manual control potentially indicating that a driver is not closely monitoring the automated system.

Predicting takeover response to silent automated vehicle failures

Current and foreseeable automated vehicles are not able to respond appropriately in all circumstances and require human monitoring. An experimental examination of steering automation failure shows that response latency, variability and corrective manoeuvring systematically depend on failure severity and the cognitive load of the driver. The results are formalised into a probabilistic predictive model of response latencies that accounts for failure severity, cognitive load and variability within and between drivers. The model predicts high rates of unsafe outcomes in plausible automation failure scenarios. These findings underline that understanding variability in failure responses is crucial for understanding outcomes in automation failures.

How Do Drivers Respond to Silent Automation Failures? Driving Simulator Study and Comparison of Computational Driver Braking Models

OBJECTIVE

This paper aims to describe and test novel computational driver models, predicting drivers' brake reaction times (BRTs) to different levels of lead vehicle braking, during driving with cruise control (CC) and during silent failures of adaptive cruise control (ACC).

BACKGROUND

Validated computational models predicting BRTs to silent failures of automation are lacking but are important for assessing the safety benefits of automated driving.

METHOD

Two alternative models of driver response to silent ACC failures are proposed: a looming prediction model, assuming that drivers embody a generative model of ACC, and a lower gain model, assuming that drivers' arousal decreases due to monitoring of the automated system. Predictions of BRTs issued by the models were tested using a driving simulator study.

RESULTS

The driving simulator study confirmed the predictions of the models: (a) BRTs were significantly shorter with an increase in kinematic criticality, both during driving with CC and during driving with ACC; (b) BRTs were significantly delayed when driving with ACC compared with driving with CC. However, the predicted BRTs were longer than the ones observed, entailing a fitting of the models to the data from the study.

CONCLUSION

Both the looming prediction model and the lower gain model predict well the BRTs for the ACC driving condition. However, the looming prediction model has the advantage of being able to predict average BRTs using the exact same parameters as the model fitted to the CC driving data.

APPLICATION

Knowledge resulting from this research can be helpful for assessing the safety benefits of automated driving.

Getting Back Into the Loop: The Perceptual-Motor Determinants of Successful Transitions out of Automated Driving

OBJECTIVE

To present a structured, narrative review highlighting research into human perceptual-motor coordination that can be applied to automated vehicle (AV)-human transitions.

BACKGROUND

Manual control of vehicles is made possible by the coordination of perceptual-motor behaviors (gaze and steering actions), where active feedback loops enable drivers to respond rapidly to ever-changing environments. AVs will change the nature of driving to periods of monitoring followed by the human driver taking over manual control. The impact of this change is currently poorly understood.

METHOD

We outline an explanatory framework for understanding control transitions based on models of human steering control. This framework can be summarized as a perceptual-motor loop that requires (a) calibration and (b) gaze and steering coordination. A review of the current experimental literature on transitions is presented in the light of this framework.

RESULTS

The success of transitions are often measured using reaction times, however, the perceptual-motor mechanisms underpinning steering quality remain relatively unexplored.

CONCLUSION

Modeling the coordination of gaze and steering and the calibration of perceptual-motor control will be crucial to ensure safe and successful transitions out of automated driving.

APPLICATION

This conclusion poses a challenge for future research on AV-human transitions. Future studies need to provide an understanding of human behavior that will be sufficient to capture the essential characteristics of drivers reengaging control of their vehicle. The proposed framework can provide a guide for investigating specific components of human control of steering and potential routes to improving manual control recovery.

Effects of an Active Visuomotor Steering Task on Covert Attention

In complex dynamic tasks such as driving it is essential to be aware of potentially important targets in peripheral vision. While eye tracking methods in various driving tasks have provided much information about drivers’ gaze strategies, these methods only inform about overt attention and provide limited grounds to assess hypotheses concerning covert attention. We adapted the Posner cue paradigm to a dynamic steering task in a driving simulator. The participants were instructed to report the presence of peripheral targets while their gaze was fixed to the road. We aimed to see whether and how the active steering task and complex visual stimulus might affect directing covert attention to the visual periphery. In a control condition, the detection task was performed without a visual scene and active steering. Detection performance in bends was better in the control task compared to corresponding performance in the steering task, indicating that active steering and the complex visual scene affected the ability to distribute covert attention. Lower targets were discriminated slower than targets at the level of the fixation circle in both conditions. We did not observe higher discriminability for on-road targets. The results may be accounted for by either bottom-up optic flow biasing of attention, or top-down saccade planning.

Toward Computational Simulations of Behavior During Automated Driving Takeovers: A Review of the Empirical and Modeling Literatures

OBJECTIVE

This article provides a review of empirical studies of automated vehicle takeovers and driver modeling to identify influential factors and their impacts on takeover performance and suggest driver models that can capture them.

BACKGROUND

Significant safety issues remain in automated-to-manual transitions of vehicle control. Developing models and computer simulations of automated vehicle control transitions may help designers mitigate these issues, but only if accurate models are used. Selecting accurate models requires estimating the impact of factors that influence takeovers.

METHOD

Articles describing automated vehicle takeovers or driver modeling research were identified through a systematic approach. Inclusion criteria were used to identify relevant studies and models of braking, steering, and the complete takeover process for further review.

RESULTS

The reviewed studies on automated vehicle takeovers identified several factors that significantly influence takeover time and post-takeover control. Drivers were found to respond similarly between manual emergencies and automated takeovers, albeit with a delay. The findings suggest that existing braking and steering models for manual driving may be applicable to modeling automated vehicle takeovers.

CONCLUSION

Time budget, repeated exposure to takeovers, silent failures, and handheld secondary tasks significantly influence takeover time. These factors in addition to takeover request modality, driving environment, non-handheld secondary tasks, level of automation, trust, fatigue, and alcohol significantly impact post-takeover control. Models that capture these effects through evidence accumulation were identified as promising directions for future work.

APPLICATION

Stakeholders interested in driver behavior during automated vehicle takeovers may use this article to identify starting points for their work.

A computational model for driver's cognitive state, visual perception and intermittent attention in a distracted car following task

We present a computational model of intermittent visual sampling and locomotor control in a simple yet representative task of a car driver following another vehicle. The model has a number of features that take it beyond the current state of the art in modelling natural tasks, and driving in particular. First, unlike most control theoretical models in vision science and engineering-where control is directly based on observable (optical) variables-actions are based on a temporally enduring internal representation. Second, unlike the more sophisticated engineering driver models based on internal representations, our model explicitly aims to be psychologically plausible, in particular in modelling perceptual processes and their limitations. Third, unlike most psychological models, it is implemented as an actual simulation model capable of full task performance (visual sampling and longitudinal control). The model is developed and validated using a dataset from a simplified car-following experiment (N = 40, in both three-dimensional virtual reality and a real instrumented vehicle). The results replicate our previously reported connection between time headway and visual attention. The model reproduces this connection and predicts that it emerges from control of action uncertainty. Implications for traffic psychological models and future developments for psychologically plausible yet computationally rigorous models of full natural task performance are discussed.

Modelling how drivers respond to a bicyclist crossing their path at an intersection: How do test track and driving simulator compare?

Bicyclist fatalities are a great concern in the European Union. Most of them are due to crashes between motorized vehicles and bicyclists at unsignalised intersections. Different countermeasures are currently being developed and implemented in order to save lives. One type of countermeasure, active safety systems, requires a deep understanding of driver behaviour to be effective without being annoying. The current study provides new knowledge about driver behaviour which can inform assessment programmes for active safety systems such as Euro NCAP. This study investigated how drivers responded to bicyclists crossing their path at an intersection. The influences of car speed and cyclist speed on the driver response process were assessed for three different crossing configurations. The same experimental protocol was tested in a fixed-base driving simulator and on a test track. A virtual model of the test track was used in the driving simulator to keep the protocol as consistent as possible across testing environments. Results show that neither car speed nor bicycle speed directly influenced the response process. The crossing configuration did not directly influence the braking response process either, but it did influence the strategy chosen by the drivers to approach the intersection. The point in time when the bicycle became visible (which depended on the car speed, the bicycle speed, and the crossing configuration) and the crossing configuration alone had the largest effects on the driver response process. Dissimilarities between test-track and driving-simulator studies were found; however, there were also interesting similarities, especially in relation to the driver braking behaviour. Drivers followed the same strategy to initiate braking, independent of the test environment. On the other hand, the test environment affected participants' strategies for releasing the gas pedal and regulating deceleration. Finally, a mathematical model, based on both experiments, is proposed to characterize driver braking behaviour in response to bicyclists crossing at intersections. This model has direct implications on what variables an in-vehicle safety system should consider and how tests in evaluation programs should be designed.