Key issues

• coordinating vehicles $\rightarrow$ decision making process orchestrating vehicles' actions
• necessary to achieve own and shared ones
• competitive/collaborative, resource/task -oriented
• poses 3 challenges
• safety: no collisions, usually
• liveness: problem specific*
• quality: problem specific*

*e.g. intersection crossing $\rightarrow$ no waiting forever, max throughput

Problems overview

Autonomy-based categorisation

Approaches categorised based on degree of autonomy left to vehicles during coordination

i.e. to what extent vehicles can decide their own actions

Categories of approaches

centralised
• burden of coordination onto a single entity (Intersection Manager)
• every vehicles abides to IM decisions by design
negotiation-based
• burden distributed amongst vehicles (and possibly a broker)
• fixed coordination protocol, fixed goal, fixed strategies
agreement-based
• burden distributed amongst vehicles
• dynamic coordination protocol, dynamic goal, switching strategies
emergent
• burden distributed amongst vehicles
• no explicit protocol, no shared goals

Intersection crossing: literature

centralised: the IM decides
• receives movement parameters from approaching vehicles
• elaborates collision-free trajectories
• communicates back to vehicles their new parameters for crossing
• e.g. constrained optimisation, coop. cruise control, reservation-based
negotiation: competitive problem $\rightarrow$ auctions
• vehicles bid for space-time slots within the intersection
• non-colliding vehicles cross simultaneously
• a broker collects bids and ranks them to assign the right of way
• many variations depending on kind of auction (e.g. English vs. Dutch), bidding strategy, ranking policy, ...
agreement: (un)structured interactions
• e.g. DCOP, argumentation-based coordination, repeated coordination games, ...
emergent: no explicit coordination protocol
• e.g. game-theoretic approaches, self-organisation

Intersection crossing: discussion

centralised
• easiest to rigorously engineer and formally verify
• guarantees of safety and liveness often available
• the IM is a bottleneck, and demands for dedicated infrastructure
negotiation-based
• still formally verifiable as protocol is fixed
• can still guarantee safety if vehicles comply (no attackers)
• liveness can be an issue especially for auctions
• neutrality issues may arise (the rich always win)
agreement-based
• not much literature in this category
• usually difficult to implement, analyse, and control
• but may provide several benefits, e.g. flexibility and explainability
emergent
• difficult to guarantee any property
• potentially relevant in mixed scenarios (i.e. with human drivers too)

Why adjustable autonomy

A single approach can hardly fit all possible run-time situations

low traffic $\rightarrow$ negotiation/agreement based
• more autonomy to vehicles
• still manageable in number of messages
high traffic $\rightarrow$ centralised approaches
• less autonomy to vehicles
• strict control to avoid collisions and delays

Thus: switch 'em!

How adjustable autonomy

• Open issue in cooperative driving, proposals in robotics and MAS
• Several challenges (akin to MAPE-K loop)
• learn best approach for each problem and specific situation
• have functional and non-functional performance metrics available to decide
• decide when to switch, possibly before degradation of efficacy
• decide how to switch (can't "pause" vehicles)

Further challenges

Besides adjustable autonomy, cooperative coordinated driving still in its infancy

• system of systems coordination (e.g. "butterfly effects" in intersection networks)
• dynamic pricing mechanisms may arise, threatening fairness and road neutrality
• dealing with mixed scenarios, as level 5 self-driving will not replace legacy vehicles overnight