Causal-based approaches to explain and learn from self-extension - A review

Rebeca Marfil; Pablo Bustos; Antonio Bandera

cs.ROOct 23, 2026

Causal-based approaches to explain and learn from self-extension - A review

Authors: Rebeca Marfil, Pablo Bustos, Antonio Bandera

Organizations: Universidad de M´alaga, Spain · Universidad de Extremadura, Spain

Abstract

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.

Figures & tables

Figure 1: A CLARA robot, equipped with a D435i camera from Intel, a pair of RFID antennas and a face recognition devoted hardware (Intel F455). The robot uses a SCITOS G3 as the base platform [ 71 ]

Figure 2: The three rungs in the Causality ladder (adapted from [ 65 ] )

Approach	Knowledge representation	Step (Figure 2 )	Application
Gao et al. [24]	And-Or graph	Intervention	Human-robot collaboration tasks
Han et al. [34]	Behaviour Trees	Intervention	Explainable planning
Lindner and Olz [53]	PDDL domain and E/D (causal) links	Intervention	Explainable planning
Krarup et al. [49]	extended PDDL2.1	Counterfactuals	Explainable planning
Diehl and Ramirez-Amaro [19]	Bayesian networks	Intervention	Failure explanation
Stange et al. [84]	A collection of memories and models	Intervention	Human-robot interaction

Table 1: Knowledge representation in autonomous robotics and their relationships with explainability

Figure 3: The hierarchical mind model for a collaboration task represented by a spatial-temporal-causal And-Or graph [ 24 ] . The task ”Picking up a trolley” implies that the robot must detect and approach to the trolley. Once the robot is correctly positioned close to the trolley, it can either ask the person to manually place the trolley on its pallet truck (because, for example, space constraints prevent the robot from doing so) or it can move and pick it up itself.

Figure 4: Schematic view of the counterfactual explanation generation in the Gjærum et al. [28] ’s proposal (see text for details)

Figure 5: Person and robot are internal plans $\overline{\pi}^{R}_{R}$ and $\overline{\pi}^{H}_{H}$ , respectively, generated with respect to their internal models. The person has no information about the robot’s model. Or the robot is not explaining (or correctly explaining) its plan to the person. In any case, the problem is that the robot’s plan is interpreted by the person ( ${\pi}_{R}^{H}$ ) with respect to his model ${\Pi}^{H}_{H}$ [ 13 ]

Requirement	Solutions
The sequential decision-making $\Pi^{R}_{R}$ must be interpretable by the person. That is, given $\mathcal{A}^{R}_{R}$ and ${\tau}^{R}_{R}$ ${\mathcal{A}}^{R}_{R}:{\Pi}^{R}_{R}\times{\tau}^{R}_{R}\nrightarrow\emptyset$ (9) When $\pi^{R}_{R}$ is derived using these and ${\Pi}^{R}_{R}$ , without encountering an error, the decision-making space is defined as interpretable.	$\Pi_{R}^{R}$ must be composed of states and actions natural to people, with transitions between states that are consistent with the real world. One solution is to manually encode the model using a collection of rules, but it is obviously complex that this set will be complete [ 74 ] . The solution can be to build a causal graph [ 55 , 19 ] . For instance, Stocking et al. [85] analyse how to distinguish task-relevant and -irrelevant variables exploring causality. Brawer et al. [6] propose an approach for a robot to learn an explicit model of cause-and-effect by constructing a structural causal model. Diehl and Ramirez-Amaro [19] describe an approach for learning a causal Bayesian network from simulated task executions. The obtained model was used for transferring the acquired knowledge from simulation to several real robots with different embodiments.
The robot must possess the tools ( $\Pi^{R}_{H}$ , $\mathcal{A}^{R}_{H}$ , and $\tau^{R}_{H}$ ) that allow it to infer the person’s behaviour. ${\mathcal{A}}^{R}_{H}:{\Pi}^{R}_{H}\times{\tau}^{R}_{H}\rightarrow\pi^{R}_{H}$ (10) where $\pi^{R}_{H}$ is a close estimate of $\pi^{H}_{H}$ .	The Bayesian framework has been extensively used to predict human reasoning in goal-oriented tasks [ 43 , 42 ] . Gao et al. [24] propose to understand complex human activities using an action parsing algorithm based on a graph-based task representation. This graph allows the robot to infer human mental states in complex environments.
The robot must be able to generate explanations that allow the person the comprehension of the robot’s plan $\pi_{R}^{R}$ . This explanation should allow the person to have an estimate, as close as possible, to the plan executed by the robot, but not necessarily to understand the robot’s policy.	Explanations must achieve a real interpretable decision-making space. In Figure 5 the situation has been exemplified in a very drastic way: the model that the person uses to interpret the robot’s plan is not influenced by the robot itself. Assuming that there is some influence (and that the model of the robot’s behaviour is estimated in some way by the person, $\Pi_{R}^{H}$ ), and denoting by $\epsilon$ the explanation provided by the robot, the objective is that $\Pi_{R}^{H}+\epsilon$ allows the person to internalize a model $\widehat{\Pi}_{R}^{H}$ such that $\mathcal{A}_{R}^{H}:\widehat{\Pi}_{R}^{H}\times\tau_{R}^{H}\rightarrow\pi_{R}^{R}.$ (11) Different approaches have been proposed to generate explanations in the field of human-robot interaction. Typically, these works seek to provide explanations in a particular order or to interweave them with the execution of the plan. Since the person does not always want the same level of detail in the explanation, Zakershahrak and Ghodratnama [94] propose a hierarchical model of explanations and details that the robot requires to provide explanations as a target. Han et al. [34] propose to encode high-level robot explanations using Behaviour Trees (BTs), specifically giving answers to questions seeking causal information. In the proposal by Gao et al. [24] , the internal model of the person is inferred and compared with the robot’s model. Whenever a disparity between these two models is detected, an explanation is generated to encourage a correction of the person’s internal state.
Explanations are not correctly encoded/conveyed to the person	We must note that, solving Equation ( 11 ), is only the first step for providing an explanation. That is, after determining the explanation $\epsilon$ , the robot needs to encode it into a form that can be conveyed to people. This problem can be solved using verbalization [ 37 ] or visualization [ 21 ] schemes. But, in general, the goal should be to personalise the communication channel according to the preferences of the end-users. In [ 72 ] , the human-robot interaction employs several informative channels and this personalisation is achieved by using an adaptation framework. It is also relevant how the information is presented to the human. Most work on explanation generation assumes that the person understands an explanation, regardless of how much information it includes. As Zakershahrak et al. [95] pointed out, complex explanations should be provided in an online fashion, broken each explanation into multiple parts, which are then provided to the person separately and intertwined with the execution of the plan.

Table 2: Issues for achieving an explainable robot [ 74 ]

Figure 6: Perceptual anomaly and the search for a causal explanation

Figure 7: KnowRobSYM software architecture (see Haidu et al. [33] for details)

Figure 8: Number of papers including the Topics Explanation and Robot. Web of Science Analyze filter: TS=(explanation) AND TS=(robot) ( https://clarivate.com/webofsciencegroup/solutions/web-of-science/ , Accessed on January 13, 2024)

Self-Evolving Agents

Self-evolving agents improve their skills, memory, tools, policies, or environments by learning from interaction and evaluation feedback.

See all 416 papers

Evidence of an Emergent "Self" in Continual Robot Learning
Mar 25, 2026Adidev Jhunjhunwala, Judah Goldfeder, Hod LipsonContinual Robot Learning
Creative Machines Lab, Department of Mechanical Engineering, Columbia University, New York, NY · Creative Machines Lab, Department of Computer Science, Columbia University, New York, NY
Self-Evolving AI for Humanoids: Mechanisms, Safety, and Evaluation of Post-Deployment Self-Improvement
Sep 2, 2026Loc X. Nguyen, Avi Deb Raha, Huy Q. Le +3Self-Evolving Agents Continual Robot Learning
School of Computing, Kyung Hee University, Yongin-si, 17104, Republic of Korea · G-LAMP NEXUS Institute, Kyung Hee University, Yongin-si, 17104, Republic of Korea · College of Computing and Data Science, Nanyang Technological University, Singapore
Ontology-Guided Reasoning for Affordance-Based Explanations of Robot Navigation
May 27, 2026Amar Halilovic, Vahidin Hasic, Senka KrivicOntology-Based Reasoning Robot Navigation
Institute of Artificial Intelligence, Ulm University, 89081 Ulm, Germany · Faculty of Electrical Engineering, University of Sarajevo, 71000 Sarajevo, BiH

Mar 25, 2026J/K move · Enter open · S save

Evidence of an Emergent "Self" in Continual Robot Learning

Adidev Jhunjhunwala, Judah Goldfeder, Hod Lipson

Creative Machines Lab, Department of Mechanical Engineering, Columbia University, New York, NY · Creative Machines Lab, Department of Computer Science, Columbia University, New York, NY

Continual Robot Learning

Open paper AskPDF arXiv

A key challenge to understanding self-awareness has been a principled way of quantifying whether an intelligent system has a concept of a "self", and if so how to differentiate the "self" from other cognitive structures. We propose that the "self" can be isolated by seeking the invariant portion of cognitive process that changes relatively little compared to more rapidly acquired cognitive skills - because our self is the most persistent aspect of our experiences. We used this principle to analyze the cognitive structure of robots under two conditions: One robot learns a constant task, while a second undergoes continual learning under variable tasks. We find that robots subjected to continual learning develop an invariant subnetwork that is significantly more stable (p < 0.001) compared to the control, and that this subnetwork is also functionally important: preserving it aids adaptation while damaging it impairs performance. We validate this pattern across three different robots spanning locomotion and manipulation.

Abstract

Self-Evolving Agents

Explore similar work

Evidence of an Emergent "Self" in Continual Robot Learning

Self-Evolving AI for Humanoids: Mechanisms, Safety, and Evaluation of Post-Deployment Self-Improvement

Ontology-Guided Reasoning for Affordance-Based Explanations of Robot Navigation