Simulation-Based Counterfactual Causal Discovery on Real World Driver Behaviour

TL;DR

The paper tackles the challenge of learning causal links between driving agents in real-world, non-stationary environments where interventions are unsafe. It introduces a simulation-based counterfactual causal discovery framework built on an event-based representation of agent decisions and a theory-of-mind model to connect base, goal, and actuation variables. Three variants—reward-based, agency-based, and hybrid—are developed and evaluated against observational temporal CD methods on 3396 scenes from the HighD dataset, showing significant performance gains, particularly for the agency-based and hybrid approaches, with runtimes compatible with online use. The work demonstrates that counterfactual simulation can yield informative causal insights beyond observational or purely intervention-based approaches, enabling safer, more interpretable real-time reasoning and retrospective analysis in autonomous driving.

Abstract

Being able to reason about how one's behaviour can affect the behaviour of others is a core skill required of intelligent driving agents. Despite this, the state of the art struggles to meet the need of agents to discover causal links between themselves and others. Observational approaches struggle because of the non-stationarity of causal links in dynamic environments, and the sparsity of causal interactions while requiring the approaches to work in an online fashion. Meanwhile interventional approaches are impractical as a vehicle cannot experiment with its actions on a public road. To counter the issue of non-stationarity we reformulate the problem in terms of extracted events, while the previously mentioned restriction upon interventions can be overcome with the use of counterfactual simulation. We present three variants of the proposed counterfactual causal discovery method and evaluate these against state of the art observational temporal causal discovery methods across 3396 causal scenes extracted from a real world driving dataset. We find that the proposed method significantly outperforms the state of the art on the proposed task quantitatively and can offer additional insights by comparing the outcome of an alternate series of decisions in a way that observational and interventional approaches cannot.

Figures (4)

• Figure 1: Illustrates a 3D depiction of the two vehicle convoy scenario considered in this paper, as well as a graph of example vehicle velocities over time for such a scenario, and the entity-level ground truth causal graph. Colour is consistent for each agent across elements.
• Figure 2: Illustration of the theory of mind ascribed to the way agents interact with the world. Circles represent data, whereas squares represent processes.
• Figure 3: Illustrates the outcomes associated with the four sets of decisions considered for a single causal link test from the head convoy vehicle to the tail convoy vehicle. The convoy vehicles are indicated in the visualisation by the lack of black rims on the central circles of the agents. Note that the convoy tail only experiences a collision if $\mathcal{C}$ occurs and $\mathcal{E}$ does not.
• Figure 4: Results of applying the three variants of the counterfactual method and baselines to the 3396 causal scenes extracted from the High-D dataset.