TL;DR
This paper introduces a novel human-machine shared control approach for Cooperative Adaptive Cruise Control (CACC) takeovers, modeling the interaction as a Stackelberg game where the machine leads. This approach ensures smoother transitions, enhances safety, and maintains platoon stability by integrating human driver behavior into the machine's predictive control, outperforming direct human takeover.
What Problem Does This Paper Address?
Cooperative Adaptive Cruise Control (CACC) offers significant benefits like increased road capacity and energy efficiency. However, a critical challenge arises during takeovers, where human drivers must regain control from the automated system. CACC's close-following strategy can make drivers feel unsafe, leading to abrupt actions like hard braking, which can cause collisions. Existing methods for automated driving takeovers often struggle with the specific dynamics of CACC, particularly the need for smooth longitudinal control transitions. This paper tackles the problem of ensuring safe, smooth, and stable CACC takeovers by developing a shared control framework that harmonizes machine and human actions.
Methods
The core of the proposed solution is an indirect human-machine shared control approach modeled as a Stackelberg competition. Here, the machine acts as the leader, and the human driver is the follower.
- Human Modeling: A Model Predictive Control (MPC) based approach is used to model the human driver's reaction function. This function captures how a human driver is likely to respond to the environment and the machine's actions.
- Machine Controller (GMPC): The machine controller is formulated as a Game-based Model Predictive Control (GMPC) problem. Crucially, the human reaction function is integrated into the machine's shared system dynamics model. This allows the machine to proactively plan its actions, anticipating and guiding the human driver's responses to achieve a stable and smooth takeover.
- Stackelberg Competition: By framing the interaction as a Stackelberg game, the machine can optimize its control strategy to influence the human's behavior, ensuring that the human's actions align with the machine's expectations for stability and safety.
- Indirect Control Fusing: A control fusing module combines the human and machine commands. It dynamically allocates authority based on factors like the Operational Design Domain (ODD) for string stability and the human driver's current actions, prioritizing smooth transitions and safety.
- Inverse Reinforcement Learning (IRL): An IRL-based method is used for driving style identification, allowing the system to adapt to different human driver behaviors by estimating weighting factors for cost functions.
Datasets
Not discussed in the paper. The evaluation relies on simulations.
Key Insights and Ideas
- Machine as a Leader: The most significant insight is modeling the machine as a leader in the takeover process. Instead of passively reacting to human actions, the machine actively guides the human, leading to more predictable and stable transitions.
- Integrating Human Model into Planning: Moving beyond a simple "predict-plan" pipeline, integrating the human reaction function directly into the machine's predictive control system is key to achieving planning optimality and proactive control.
- Stackelberg Game for Takeover: Applying the Stackelberg game framework effectively models the hierarchical control authority during a takeover, where the machine's objective is to influence the human's optimal response.
- ODD for String Stability: Quantifying the Operational Design Domain (ODD) for string stability based on authority allocation is crucial for practical implementation, revealing that human authority needs to be limited for stable platoons.
- Trade-off Between Space and Time: The proposed method demonstrates a good trade-off between the duration of the takeover and the overall time to stabilize the platoon, offering flexibility in adapting to different scenarios.
Key Results
- Smooth Takeover: The proposed controller enables smooth CACC takeover maneuvers, as evidenced by reduced speed oscillation magnitude and smoother acceleration profiles compared to direct human takeover.
- String Stability: String stability is ensured when the human control authority is less than 40% (specifically, less than 28.6% for 10 CAVs) and the platoon length is restricted (ideally under 6 CAVs).
- Enhanced Safety: Machine interventions significantly enhance perceived safety (by at least 53.23%) and actual safety, preventing hard braking and maintaining safe following gaps, especially during hard braking scenarios by the leading vehicle.
- Reduced Traffic Influence: The controller reduces traffic oscillation propagation by up to 60%, with the machine's greater control authority leading to fewer oscillations.
- Driving Style Identification: The IRL-based driving style identification method achieves an accuracy of 91.6%.
- Computational Efficiency: The proposed controller demonstrates efficient computation, with an average computing time of 0.05s in a distributed framework.
Limitations and Future Work
- ODD Limitations: The paper quantifies string stability ODDs, highlighting that stability deteriorates with increased human authority and platoon length. This suggests practical limitations on platoon size and human intervention levels.
- Automation-Initiated Focus: The current work primarily focuses on automation-initiated takeover transitions.
- Future Work: The authors suggest exploring adaptive arbitration strategies for driver-initiated transitions and incorporating empirical studies of driver behavior during voluntary overrides. Further research could also investigate harmonizing driver intent with automation trajectories.
Summary updated