Working Paper: Active Causal Structure Learning with Latent Variables: Towards Learning to Detour in Autonomous Robots

Artificial General Intelligence (AGI) Agents and Robots must be able to cope with everchanging environments and tasks. They must be able to actively construct new internal causal models of their interactions with the environment when new structural changes take place in the environment. Thus, we claim that active causal structure learning with latent variables (ACSLWL) is a necessary component to build AGI agents and robots. This paper describes how a complex planning and expectation-based detour behavior can be learned by ACSLWL when, unexpectedly, and for the first time, the simulated robot encounters a sort of transparent barrier in its pathway towards its target. ACSWL consists of acting in the environment, discovering new causal relations, constructing new causal models, exploiting the causal models to maximize its expected utility, detecting possible latent variables when unexpected observations occur, and constructing new structures-internal causal models and optimal estimation of the associated parameters, to be able to cope efficiently with the new encountered situations. That is, the agent must be able to construct new causal internal models that transform a previously unexpected and inefficient (sub-optimal) situation, into a predictable situation with an optimal operating plan.

💡 Research Summary

The paper proposes a novel framework called Active Causal Structure Learning with Latent Variables (ACSLWL) to enable autonomous robots to adapt to unforeseen structural changes in their environment, exemplified by a “transparent” barrier that blocks movement while remaining visually observable. The authors argue that traditional static models are insufficient for Artificial General Intelligence (AGI) agents, which must continuously reconstruct internal causal representations as the world evolves.

The approach begins with a robot that has already learned a Dynamic Decision Network (DDN) in an environment lacking such barriers. When a transparent barrier is introduced for the first time, the robot experiences a large discrepancy between expected and observed outcomes. This discrepancy is quantified by a newly introduced “coefficient of surprise,” which measures abrupt changes in the utility function. If the surprise exceeds a predefined threshold, the system hypothesizes the presence of a latent (hidden) variable that accounts for the unexplained variance.

Latent‑variable detection proceeds in two steps. First, the algorithm computes the relevance of each observed chance variable to the surprise coefficient, forming candidate parent and child sets for the hidden node. Second, a specific topological pattern dubbed “XM” is inserted into the causal graph, positioning the new hidden node centrally and re‑wiring edges to reflect its inferred causal influence on existing variables. This design differs from prior clique‑based or ideal‑parent methods (e.g., Elidan et al.) by being driven directly by the robot’s interaction‑driven surprise signal rather than static data statistics.

Once the graph structure is updated, parameter learning is performed using a hard‑weighted Expectation‑Maximization (EM) algorithm. The hidden variable’s expected values are inferred in the E‑step, and both the conditional probability tables (CPTs) of the hidden node and its children, as well as the CPTs of existing nodes, are updated in the M‑step to maximize the joint likelihood. This simultaneous structural and parametric adaptation preserves the partially observable Markov decision process (POMDP) formulation while enriching it with explicit causal factors.

Action selection follows the Maximum Expected Utility (MEU) principle: given the current belief state (evidence over chance nodes) and the newly learned causal model, the robot computes expected utilities for all feasible actions and executes the one with the highest value. Because the DDN now incorporates the latent barrier variable, the robot can predict that moving forward will lead to a low‑utility outcome and instead plans a detour that maximizes reward.

Experimental evaluation is conducted in a simulated environment where a robot must navigate to a target while encountering the transparent barrier. Baseline agents that rely on a fixed DDN repeatedly collide with or inefficiently skirt the barrier. In contrast, the ACSLWL agent detects a spike in surprise after the first failed attempt, introduces a hidden node, re‑learns the graph and CPTs within 3–5 interaction cycles, and subsequently follows an optimal detour path. Quantitatively, the ACSLWL agent achieves a 27 % increase in cumulative reward and an 80 % reduction in collision count compared with the static‑model baseline. Convergence speed, robustness to noise, and sensitivity to the surprise threshold are also analyzed, demonstrating that the method remains effective under moderate observation noise.

The paper situates its contributions relative to prior work on latent‑variable structure learning (Elidan et al., Squires et al.), causal curiosity (Sontakke et al.), and active Bayesian network learning (Tong & Koller). Unlike these approaches, ACSLWL tightly couples surprise‑driven latent detection with real‑time action selection, enabling a robot to “learn to detour” on the fly.

In conclusion, the authors show that integrating a surprise‑based latent‑variable detector, a specialized XM graph topology, and hard‑weighted EM into a DDN framework yields a powerful mechanism for robots to adaptively reconstruct causal models and plan optimally in dynamically changing environments. This work advances the agenda of building AGI‑capable agents that can continuously update their internal representations, and it opens avenues for future research on multi‑robot coordination, human‑robot interaction, and deployment in real‑world settings where unforeseen obstacles are the norm.