Continual Causal Discovery for Robotics Applications

Learning the causal relationships underlying the dynamics of a shared environment from time-series of sensor data is a key problem for robots operating alongside humans. However, in non-stationary environments, batch-estimated models silently cease to describe the system, while continuously repeating causal discovery is computationally prohibitive for resource-constrained autonomous platforms. In this thesis, we propose CoRe-SCD (Constraint Replay for Structural Change Detection), an online method that detects when a causal model is no longer valid by monitoring the conditional independence constraints established during initial discovery, reporting only changes that alter the estimated graph. We validate the approach on synthetic structural causal models and on continuous physical data from a simulated ROS/Gazebo human-robot interaction scenario. Experiments demonstrate that CoRe-SCD detects structural modifications and explicitly names the responsible constraints. Crucially, by monitoring unstandardised regression coefficients rather than standard p-values, the method mathematically decouples causal effect strength from measurement precision. This enables the system to inherently absorb non-structural parametric shifts and sensor noise, preventing spurious topological alarms. Ultimately, CoRe-SCD provides an inexpensive, robust criterion to determine when full causal re-discovery is genuinely warranted, supporting adaptive causal modelling for safe, long-term robot autonomy.