Intent Communication between Autonomous Vehicles and Pedestrians

I. INTRODUCTION

• The paper addresses a gap in quantitative experiments analyzing ICS utility in real vehicle-pedestrian interactions.It also examines how pedestrians perceive autonomous vehicles and what they expect from machines.Source: §I · pp. 1–2
• The paper contributes an ICS designed to diffuse vehicle-pedestrian deadlock, a mathematical model of how vehicle decisions affect pedestrian trust, and evaluation through experiments.The authors describe the evaluation as real-world testing and simulations.Source: §I · pp. 1–2

II. RELATED WORK

• The authors argue that careful human-feedback evaluation matters for usable intent communication, noting that Google’s patent discussion did not reveal testing, robustness, or cost effectiveness,.They call for further research on pedestrian psychology, receptivity, trust, and safety.Source: §II · p. 2
• Prior human-machine interaction work reports greater receptivity to robotic instructions when dialogue is present, while this paper adds quantitative trust measurement in pedestrian-vehicle encounters.The paper uses post-surveys to measure perceived effects on trust.Source: §II · p. 2
• The related work discusses trust modeling with MDP variants, including Dec-MDPs, to account for stochastic human actions,,.This paper uses the Dec-MDP framework for simulations.Source: §II · p. 2
• Prior approaches include gesture and audio interfaces, but gestures can be difficult for machines to interpret and audio cues depend on context,,,,.The paper presents its approach as combining trust with implicit and explicit communication for people without prior machine interaction.Source: §II · p. 2

A. Intent Communication System Description

• The ICS combines hardware and software, using strobes to attract attention and an LED word display and speakers to communicate what the golf car wants pedestrians to do.A camera detects pedestrians, and software signals the display, speakers, and lights.Source: A. Intent Communication System… · p. 3
• The design aims to avoid intimidating pedestrians and accounts for durability, cost, and operation in varied environments.The authors cite concern that more human-like machines can unsettle people.Source: A. Intent Communication System… · p. 3

B. Intent Communication Psychology

• The authors prioritize how people perceive the ICS, using familiar road cues such as words, flashing lights, and sounds rather than complicated communication.They report that initial user feedback shaped this design consideration.Source: B. Intent Communication Psychology · p. 3
• Drawing on human dialogue and trust, the ICS uses short instructions such as “please cross” to leave little room for misinterpretation.The authors argue that simple statements can guide action without in-depth conversation.Source: B. Intent Communication Psychology · pp. 3–4

C. Real World Experiment

• The golf car was driven manually by transmitter for safety, while participants were unaware of the manual control and the operator could intervene.The authors state this preserved unbiased results regarding perceived autonomy.Source: C. Real World Experiment · p. 4
• The 26 real-world testing participants were divided among four groups varying ICS enablement and prior knowledge; prior-knowledge participants were introduced to the system before testing.The full study included 76 participants, including 50 in a pre-survey, with nine involved in both sets.Source: C. Real World Experiment · p. 4
• The study varied whether the ICS was enabled and whether participants had prior knowledge, and none knew whether a human was inside the vehicle.Prior-knowledge participants were shown how the vehicle’s lidar, ICS, and other technology worked.Source: C. Real World Experiment · p. 4
• Surveys asked whether the car met expectations, whether participants behaved as expected, whether the ICS created trust, and whether participants felt safe and found communication effective.Surveys and interaction video were used to assess perceptions and effects on trust in the vehicle’s decision-making.Source: C. Real World Experiment · p. 4

D. Real World Results

• Group 1 had the highest trust ratings before and after testing, and all participants said their behavior matched their expectations.Responses were approximately split on whether the car was trying to work with people or replace humans.Source: 1) Group 1: With ICS and Prior… · pp. 4–5
• Group 2 found the car’s actions predictable and had relatively high trust, but did not trust it more than a human driver compared with groups without ICS.About 80 percent felt the car was trying to work with people rather than replace them.Source: 2) Group 2: With ICS and no Prior… · p. 5
• In Group 3, participants had prior knowledge but tested the golf car without ICS; about 60 percent found it completely unpredictable and felt somewhat unsafe, while trust was higher after the experiment.Source: 3) Group 3: Without ICS and Prior… · p. 5
• Group 4 had the lowest trust before and after testing, and about 80 percent said they behaved differently than expected.Participants also reported feeling unsafe around the car.Source: 4) Group 4: Without ICS and no Prior… · p. 5
• Groups 1 and 2, both tested with ICS, had higher overall trust than Groups 3 and 4 without ICS.Table I reports post-test survey scores scaled to a maximum of 30 points per group.Source: Table I · p. 5, 4) Group 4: Without ICS and no Prior… · p. 5
• Interaction times were 0.24 seconds for Group 1, 0.61 for Group 2, 0.39 for Group 3, and 1.01 for Group 4.Groups 1 and 3 had prior knowledge; Groups 2 and 4 did not. Groups 1 and 2 had ICS enabled.Source: 4) Group 4: Without ICS and no Prior… · p. 5, Figure 5

A. Simulation Setup

• Simulations used the same scenarios as the real-world experiment and included one vehicle and one pedestrian at a time.The paper treats pedestrian groups as one pedestrian because groups typically exhibit group behavior.Source: A. Simulation Setup · pp. 5–6
• The car’s actions included moving forward, stopping, and turning; pedestrian actions also included waiting, entering the car, and not noticing it.The pedestrian action set was derived from pre-survey responses; “wait” could continue indefinitely, whereas “stop” was temporary.Source: A. Simulation Setup · p. 6
• Pre-survey responses supplied psychological factors for pedestrian decisions, while post-survey results updated the simulated pedestrian model.The authors chose surveys to represent a broader range of possible actions without testing them by driving the car around people.Source: A. Simulation Setup · p. 6
• Trust, defined as pedestrian confidence in the vehicle, was the link between the real-world experiments, simulations, and mathematical formulation.The simulation focused on vehicle–pedestrian interaction, while also accounting for independent movement when considering trust.Source: A. Simulation Setup · p. 6

B. Simulation Methodology

• The simulation model was designed to represent stochastic pedestrian actions, changing trust, and the effectiveness of vehicle communication.The survey approach was selected instead of observing reactions by driving the autonomous vehicle because of safety concerns.Source: B. Simulation Methodology · p. 6, C. Pedestrian Surveys · p. 6

C. Pedestrian Surveys

• Fifty unaffiliated respondents supplied potential pedestrian actions and probabilities for encounters with a golf car that might have no visible operator.The survey aimed to narrow the action set and inform communication-system design.Source: C. Pedestrian Surveys · p. 6
• Respondents named actions without being shown a list, and the resulting action categories included stopping, waiting, crossing, entering the car, and not noticing it.The survey also asked how respondents would feel and what would make them more comfortable around a driverless car.Source: C. Pedestrian Surveys · p. 6

D. Dec-MDP Definition

• The model treats pedestrian intent as uncertain and uses a Dec-MDP rather than a POMDP, whose decentralized form is described as computationally intractable.The model assumes both agents can observe each other’s current physical state, although their internal states are not fully known to one another.Source: D. Dec-MDP Definition · pp. 6–7
• When ICS communicates vehicle intent, the authors propose treating the partially observable interaction as a Dec-MDP because pedestrian actions can then be reasonably predicted.The paper presents this as a way to make the solution implementable on real-world robots.Source: D. Dec-MDP Definition · p. 7
• The model assumes joint full observability: the agents’ paired observations together determine the current state.The ICS communicates the vehicle’s internal state to the pedestrian, narrowing the pedestrian’s possible actions.Source: D. Dec-MDP Definition · pp. 7, 9
• The model lets agents move independently and consider one another when they interact, with deterministic vehicle behavior and stochastic pedestrian behavior.The car is assumed controllable, and interactions are modeled over a finite horizon.Source: D. Dec-MDP Definition · pp. 7, 9
• The simulations track trust over an encounter and allow varied action sets, including potentially unsafe or abnormal pedestrian behavior.The authors describe the model as reducing computational overhead while retaining flexibility for changing scenarios.Source: D. Dec-MDP Definition · pp. 7, 9, 10
• The trust reward draws on real-world experiment results and weights interaction according to agent actions, transition probabilities, and proximity.Pedestrian transition probabilities use pre-survey action probabilities and post-survey confidence; the proximity range was set to 6 for the gridworld simulations.Source: D. Dec-MDP Definition · pp. 7–9
• The simulated pedestrian action probabilities were adjusted using survey-derived trust, with higher weights assigned to actions that were initially less likely.The example for ICS with prior knowledge reports an overall trust rating of 27 out of 30 and a 90 percent weighting.Source: D. Dec-MDP Definition · p. 9
• The finite-interval Dec-MDP lets the vehicle adapt before a pedestrian reaches sensor range, unlike a reactive strategy that may leave too little time to maneuver.The model also updates actions when pedestrian behavior is unexpected and uses proximity to set when the vehicle begins accommodating that behavior.Source: D. Dec-MDP Definition · pp. 9–10

• Each scenario uses a gridworld with one vehicle and one pedestrian, with pedestrian trust distance, action probabilities, and likelihood of following vehicle directions adjusted to match the scenario.Source: E. Simulation Results · p. 10
• Pedestrian action probabilities and following likelihood reflect real-world experiments, while trust is modeled from participant post-survey results and vehicle parameters remain unchanged.Source: E. Simulation Results · p. 10
• Pedestrians move through the gridworld as particles, moving immediately in the chosen direction rather than turning to face it.Source: E. Simulation Results · p. 10

1) With ICS and Prior Knowledge:

• The scenario modeled the pedestrian as one unit space from the vehicle.Source: 1) With ICS and Prior Knowledge · p. 10
• Trust fluctuated during encounters and was lower when the pedestrian and vehicle were closer together.Source: 1) With ICS and Prior Knowledge · p. 10
• The pedestrian took riskier actions than in the other scenarios, including getting inside the vehicle.Source: 1) With ICS and Prior Knowledge · p. 10

2) With ICS and No Prior Knowledge:

• The modeled distance from the vehicle was three unit spaces to provide a higher safety factor for the lack of prior knowledge.Source: 2) With ICS and No Prior Knowledge · p. 10
• Trust dropped drastically as soon as the vehicle started moving, compared with the prior-knowledge scenario.Source: 2) With ICS and No Prior Knowledge · p. 10
• The pedestrian sometimes took bold actions and sometimes did not notice the vehicle.Source: 2) With ICS and No Prior Knowledge · p. 10

3) Without ICS and Prior Knowledge:

• The scenario assumed pedestrians would expect commands from the vehicle, so the distance was set to one unit space.Source: 3) Without ICS and Prior Knowledge · p. 10
• The one-unit distance matched the first scenario's setting.Source: 3) Without ICS and Prior Knowledge · p. 10
• Trust was never consistent, with higher values in some pedestrian–vehicle encounters than in others.Source: 3) Without ICS and Prior Knowledge · p. 10

4) Without ICS and Prior Knowledge:

• As pedestrian confidence increased, trust values showed less variance from baseline; variance was largest with ICS and no prior knowledge and smallest without ICS and with prior knowledge.The plots group four scenarios at confidence levels of 20%, 40%, 60%, 80%, and 99%.Source: 4) Without ICS and Prior Knowledge · p. 11, Figure 9 · p. 11, Figure 10 · p. 11
• Trust remained negative in every simulated scenario, with trust gains varying across scenarios and trust changing little over 10 iterations.Each scenario averaged 2,000 runs; 10 iterations represented the time for either the vehicle or pedestrian to reach its goal.Source: 4) Without ICS and Prior Knowledge · pp. 10–11
• Prior knowledge was reported as more important than enabling the ICS, and simulations found that prior knowledge increased trust while prior knowledge or an enabled ICS was associated with shorter interaction times.The simulation’s no-prior-knowledge scenario used a maximum distance of six unit spaces.Source: 4) Without ICS and Prior Knowledge · pp. 10–12

V. CONCLUSION

• The authors report that experimentation showed the ICS could help resolve potentially dangerous and inefficient deadlock situations by 38 percent.Both real-world testing and simulations evaluated trust when pedestrians encountered an autonomous vehicle.Source: §V · p. 12
• Trust was associated with comfort around the vehicle, prior knowledge, and vehicle distance; participants more familiar with the ICS were more likely to trust it.The authors describe this as among the first autonomous-vehicle tests to confirm these factors in vehicle-human interaction.Source: §V · p. 12
• Simulations based on real-world data safely tested riskier behaviors, but trust fluctuated without leaving the negative region.Examples included pedestrians getting into the vehicle or not noticing it.Source: §V · p. 12
• Groups with prior ICS knowledge had approximately 10% and 6% higher trust than their respective comparison groups, while ICS-enabled groups had 33% and 24% higher trust than groups without ICS.Both real-world experiments and simulations also found shorter interaction times for groups with prior knowledge or an enabled ICS.Source: §V · p. 12
• In the real-world study, ICS-enabled groups were more likely to trust the vehicle but hesitated slightly before crossing; ICS-disabled groups showed less trust and considerably more hesitation.Participants sometimes viewed the ICS as a threat because they saw it as replacing human drivers.Source: §V · p. 12
• The authors identify further work on increasing trust, testing multiple pedestrians at once, and studying expectations to inform system design.Source: §V · p. 12