Source-linked AI summary
Causal-based approaches to explain and learn from self-extension - A review
Rebeca Marfil, Pablo Bustos, Antonio Bandera
TL;DR
Deep-learning robots have gained perceptual and task-solving capabilities, but their decisions can remain difficult for people to understand. This review surveys causal approaches to explaining robot actions, plan updates, and perceptual anomalies, finding that causality is increasingly treated as a basis for explanations and robot acceptability.
Problem
Method
The review examines how robots can use cause-effect relationships to explain actions, plan updates, and perceptual anomalies, then integrate explanations into knowledge for later use.
Source: §1 · pp. 2–3Results
The review finds that integrating causality into robot explanations has begun to be considered a basic pillar for robots to be seen as credible, safe, and reliable.
Source: §4 · p. 20Takeaways & Limitations
Explanations shared and validated with people can be incorporated into a robot's knowledge base and used to improve future behaviour.
Source: §1 · p. 2For shared robot-human task execution, the plan-comparison scheme does not represent plans' causal structure or address abstraction and unsolvability.
Source: §3 · p. 19
Abstract
from arXiv · showhide
The last decades have seen a revolution in autonomous robotics. Deep learning approaches and their hardware implementations have made it possible to endow robots with extraordinary perceptual capabilities. In addition, they can also benefit from advances in Automated Planning, allowing them to autonomously solve complex tasks. However, on many occasions, the robot still acts without internalising and understanding the reasons behind a perception or an action, beyond an immediate response to a current state of the context. This gap results in limitations that affect its performance, reliability and trustworthiness. Deep learning alone cannot bridge this gap, because the reasons behind behaviour, when it emanates from a model in which the world is a black-box, are not accessible. What is really needed is an underlying architecture based on deeper reasoning. Among other issues, this architecture should enable the robot to generate explanations, allowing people to know why the robot is performing, or has performed, a certain action, or the reasons that may have caused a certain plan failure or perceptual anomaly. Furthermore, when these explanations arise from a cognitive process and are shared, and thus validated, with people, the robot should be able to incorporate these explanations into its knowledge base, and thus use this understanding to improve future behaviour. Our article looks at recent advances in the development of self-aware, self-evolving robots. These robots are designed to provide the necessary explanations to their human counterparts, thereby enhancing their functional capabilities, in the quest to gain their trust.
1 Introduction
As robots enter shared daily-life settings, appropriate behaviour depends on context and on acceptability, accessibility, and adaptability. The paper frames social awareness, self-understanding, and causal explanations as central to robots that adapt to and explain their behaviour.
1 Introduction
- Service robots face especially complex technical, ethical, social, and legislative demands because they are used in daily-life settings.Appropriate behaviour also varies across societies, individuals, and contexts.Source: §1 · p. 1
- Socially-aware Robotics (SAWR) seeks cognitive robots that understand social context and behave accessibly, acceptably, and usefully.Here, a socially-aware robot proactively predicts, anticipates, evaluates, and learns from its social context to adapt its behaviour.Source: §1 · pp. 1–2
- Socially-aware robots need to recognize and predict people’s behaviour, anticipate action consequences, adapt to users, and explain their behaviour to support human trust.The paper identifies causal discovery and reasoning as a potential tool for these requirements.Source: §1 · p. 2
- The review examines how robots can causally explain plan updates and perceptual anomalies, integrate those explanations into knowledge, and use them to improve later behaviour.It also considers explanations as part of real communication with people, not only as aids to debugging or liability assessment [12].Source: §1 · p. 2
- A self-understanding system explicitly represents incompleteness in its knowledge, while a self-extending system can fill that incomplete knowledge.Source: §1 · p. 3
- The review covers unconstrained explanations of actions and perceptual anomalies, constrained plan questions, and reproducing perceptual experiences to identify anomaly causes.It treats causal models as related to planning but does not focus on how robots plan tasks.Source: §1 · p. 3
2 Climbing the ladder of causation
This section surveys causal approaches to explaining robot actions and perceptual anomalies, spanning knowledge representations, planning, and reasoning from observations to interventions and counterfactuals. It considers how explanations can be made understandable to people and how tested causal hypotheses could support revising internal models and learning.
2 Climbing the ladder of causation
- Explainable autonomous robots (XARs) explain their autonomous actions in shared environments, making their focus goal-driven rather than the data-driven explainability typical of XAI [74].The section links explanations to trust and to people understanding the robot’s internal state and intentions.Source: §2 · pp. 3–4
- Pearl and Mackenzie’s causation ladder moves from association (observing regularities), through intervention (predicting action effects), to counterfactual reasoning about why outcomes occurred and what might have happened otherwise.The section says robots are now comfortably on the association rung, while explanations at higher rungs require greater cognitive ability.Source: §2 · pp. 4–6
2.1 Knowledge representations and explainability
- Explainable-planning representations include And-Or graphs for collaborative task progress [24], Behaviour Trees for causal explanations, and PDDL models whose preconditions and postconditions support revisable, action-by-action plans and post-hoc explanations [49].PDDL also supports concurrent planning and execution; its causal models can help reconcile robot and person models.Source: §2.1 · pp. 6, 8
- Ontology-based knowledge can be reused across changing scenarios, while KnowRob supports situation emulation and planning; IID and simulator-based architectures represent state, action effects, and possible failure causes.In the simulator-based approach, inconsistencies between perceived and simulated reality trigger causal explanation and concept acquisition.Source: §2.1 · p. 9
2.2 Explaining a plan: how and why are I going to do it?
- Automated planning projects goals, preconditions, resources, and timing to predict action effects and respond to unforeseen states; explanation aims to improve objective and subjective measures such as efficiency, trust, and workload [49].The section notes that evaluation may combine task measures with Likert-scale assessments.Source: §2.2 · p. 10
- Plan explanations based only on the robot’s internal model may be useful mainly to people familiar with it; effective explanations require interpretable models, adequate robot and person models, and understandable presentation [12] [74].When the models differ, especially for non-experts, explanations of failure that include its context and past actions are most effective for identifying failures and solutions [17].Source: §2.2 · pp. 11, 12, 14
- PDDL represents planning problems declaratively, but robot control still requires mapping high-level plans to low-level sensing and reactive behaviours, typically through platform-specific code.The passage notes that declarative mappings could be edited without modifying the control architecture.Source: §2.2 · p. 14
2.3 Exploring the recent past: why did you do this instead of that?
- People often ask contrastive plan questions—“Why did you do this instead of that?”—and answering can involve iteratively constraining planning models to generate alternatives.The alternatives support comparing the person’s proposal with the original plan and can reveal the robot’s model and the person’s preferences.Source: §2.3 · pp. 15–16
- The review focuses on local contrastive questions about specific decisions or plan adequacy, and stresses that robotic counterfactual states and actions must be physically feasible.Questions can also be global, seeking general knowledge about decision-making or the model itself.Source: §2.3 · p. 15
2.4 (Self)Explaining perceptual anomalies
- For perceptual anomalies, causal explanations could prompt structural changes to a robot’s internal model, support learning new concepts or behaviours, and be rendered in natural language for a person.The transformation should account for compatibility between the human’s and robot’s models.Source: §2.4 · p. 17
- An internal simulator can detect anomalies by comparing synchronous predictions with sensor measurements, then replay the event from candidate hypotheses and compare synthetic and real sensory data.The proposed steps include converting anomaly data and hypotheses into formal representations, testing them through replay, and selecting a best hypothesis for model updates and explanation.Source: §2.4 · p. 17
- Plausible causes may come from semantic resources or an internal LLM; testing them against recorded sensory events could inform future behaviour and communication, with those actions feeding back to confirm or refute explanations.The review presents this as a promising path toward self-explaining robotic architectures while noting that open problems remain.Source: §2.4 · p. 18
3 Discussion and future research work
The review describes growing interest in autonomous robots that can explain their behaviour as robots increasingly share environments with people. It identifies open challenges in explaining plans and perceptual anomalies, including causal structure and online transfer between numerical and symbolic representations.
3 Discussion and future research work
- More than 100 publications per year in recent years versus fewer than 20 in 2000; 2023 data were incomplete.The counts concern publications on Explanation and Robot in the Web of Science search described in Figure 8.Source: §3 · p. 19, Figure 8 · p. 19
- Explanations in shared robot-human tasks often compare plans against a known or learned joint plan, and added planning constraints can be solved in similar time to the unconstrained problem.The review frames questions such as why one activity was chosen over another as restricted planning.Source: §3 · p. 19
- Future work should account for plans’ causal structure, abstraction, and unsolvability, while PDDL lacks easy expression of several action-ordering and plan-structure constraints.The review notes that trajectory-based action constraints and geometric emulation may help address some of these issues; ASP-based planners outperformed PDDL-based planners when objects and tasks were numerous and complex relationships mattered.Source: §3 · pp. 19–20
- Perceptual-anomaly explanations could use internal simulators to test contrastive hypotheses, but online transfer between numerical simulation and symbolic knowledge representations remains challenging.The review identifies transformations between these representations as a promising area for research on new concepts.Source: §3 · p. 20
4 Conclusions
The conclusion presents causal relationships as a basis for autonomous robots to generate explanations. These may concern actions in a traversed sequence or consider alternative sequences and disregarded perceptions that could explain unforeseen results.
4 Conclusions
- The article reviews a general framework for robot explanations grounded in internalized cause-effect relationships.The conclusion links this integration of causality with efforts to create robots people consider acceptable, credible, safe, and reliable.Source: §4 · p. 20
- Explanations can describe specific actions already taken or planned within a temporal sequence.Source: §4 · p. 20
- Explanations can also examine unconsidered sequences, including actions not taken or perceptions disregarded, as possible causes of unforeseen results.Source: §4 · p. 20