No Plan, Yet Human: A Reactive Robotics Model Predicts Human Planning Failures on a Clinical Task

Abstract: Understanding why some sequential planning problems are harder than others requires models that go beyond average performance. They should capture the specific pattern of which problems are hard, and ideally fail in the same way people do when planning capacity is reduced. We apply AICON, a reactive gradient-descent framework developed for robotic manipulation, to the Tower of London test, a cognitive test used to assess planning in Parkinson's disease, mild cognitive impairment, and stroke. Without any lookahead planning or knowledge of human cognition, AICON reproduces the fine-grained human difficulty ordering across 24 problems better than structural task parameters and generalizes to held-out problems in a leave-two-out evaluation. Crucially, AICON outperforms a planning baseline for groups with reduced planning capacity while the planning baseline better captures healthy controls. This dissociation was predicted by the original AICON paper, which noted that the model's failure modes resemble those of Parkinson's patients who struggle with goal hierarchies but not move counts. This suggests that as planning capacity is reduced, human behavior shifts toward the reactive mode AICON models. The finding extends a broader pattern: AICON, originally built for robotics, now captures aspects of biological behavior across perception, eye movements, and sequential planning, suggesting its core abstraction reflects something real about how biological systems are organized.

• The paper demonstrates that AICON, a reactive robotics model, replicates human planning deficits in the Tower of London task using gradient-based decision-making without explicit lookahead.
• It shows that AICON outperforms traditional planning baselines in clinical groups such as Parkinson’s, MCI, and stroke patients by aligning with empirical difficulty orderings.
• The findings support dual-process theories, suggesting that diminished planning shifts control to reactive mechanisms with potential diagnostic and therapeutic implications.

Modeling Human Planning Failures: AICON’s Predictive Power on the Tower of London Task

Introduction

This work rigorously evaluates whether a reactive robotics framework, Active InterCONnect (AICON), can mechanistically reproduce human problem-solving deficits during a classic neuropsychological planning test, the Tower of London (ToL). The authors set out to answer a pointed question: do planning-impaired humans, such as Parkinson’s patients, behave in ways that can be better explained by purely reactive—rather than deliberative—mechanisms? AICON, without any lookahead or explicit planning, is benchmarked against human data from diverse clinical groups and compared to both structural task parameters and planning-based baselines.

AICON: A Reactive Paradigm in Sequential Problem Solving

AICON departs from conventional planning models by operating in a completely reactive manner. Rather than engaging in lookahead or internal modeling of future states, AICON encodes task structure as a network of recursive estimators manipulating a differentiable state representation. The legalities and constraints of ToL are captured through active interconnections, and action selection is governed by descending the steepest gradient in a cost function reflecting current deviation from the goal configuration.

The model distinguishes legal moves and reachable subgoals entirely through real-time gradient computations, eschewing tree search or policy learning. Importantly, this architecture’s failure modes derive directly from its structure: it is competent in scenarios where immediate gradients suffice but struggles when a sequence of moves involves non-local dependencies or conflicting subgoals. This leads to a characteristic set of failures, hypothesized to mirror those of humans with diminished planning capacity.

Human Behavioral Benchmarks

Empirical validation is conducted using a dataset spanning 24 ToL problems and four groups: controls, Parkinson's disease (PD), mild cognitive impairment (MCI), and stroke patients. Human performance is measured as average additional moves beyond optimal and success rate across problem types. The principal analytic metric is Kendall’s $\tau$, evaluating the rank correlation between the model’s predicted problem difficulties and the actual empirical ordering for different groups.

Comparative Analysis: AICON Versus Planning Baselines

AICON is pitted against two baseline accounts:

• Descriptive baseline: The minimal solution length for each problem, i.e., structural task parameterization.
• Planning baseline: Breadth-first search (BFS), evaluated in unidirectional and bidirectional variants.

Bidirectional BFS, which averages forward and backward search costs, provides a competitive planning-centric account of problem difficulty, particularly among healthy subjects. However, AICON’s performance is distinguished by robust generalization in leave-two-out splits and superior predictive alignment with impaired groups.

Principal Numerical Results and Claims

• For healthy controls, bidirectional BFS matches or exceeds AICON’s $\tau$, directly supporting the conclusion that intact planning mechanisms underlie performance in that population.
• In contrast, for PD, MCI, and stroke groups, AICON produces higher $\tau$ than any planning baseline for both success rate and excess moves. This demonstrates a specific and robust advantage for AICON in capturing the nuanced difficulty ordering of problems among individuals with reduced planning capacity.
• AICON’s superiority is especially marked for PD patients and is maintained even when only “easy vs. easy” or “hard vs. hard” problem pairs are used in the evaluation, exposing the model’s sensitivity to within-level difficulty that eludes planning-centric baselines.
• Structural task parameters (optimal move counts) systematically underperform both AICON and planning baselines, emphasizing the need for mechanistic, rather than merely descriptive, models.

Theoretical Implications

This dissociation validates longstanding theoretical conjectures regarding dual-process accounts of sequential problem solving: healthy individuals exploit both reactive and reflective mechanisms, while reductions in cognitive resources (e.g., in PD) shift control toward immediate, stimulus-driven strategies. The characteristic failure modes observed in AICON—especially its insensitivity to deeper subgoal conflicts—provide a veridical computational analog of the deficits exhibited by clinical populations. This substantiates the claim that complex-looking human failures can emerge naturally from reactive architectures without invoking impaired planning as a unidimensional loss of capacity.

Additionally, the fact that AICON, initially developed for high-dimensional robotics tasks, generalizes so effectively to clinical cognitive assessment signals the relevance of such architectures for modeling real biological information processing. The authors tie these results to a broader pattern in which AICON architectures have modeled perceptual behavior, gaze allocation, and collective dynamics.

Practical and Clinical Outlook

Validated mechanistic models of ToL problem solving have practical significance for neuropsychological diagnosis and intervention strategies. AICON’s ability to predict not only group-level but item-level variance in performance may inform selection of diagnostic items, identification of cognitive phenotypes, and longitudinal tracking of progression in neurodegenerative conditions. The approach also raises the prospect of combining reactive and deliberative modules in a unified account of human sequential decision-making, explicitly modeling the balance between planning and reaction as a function of neural integrity and task structure.

Limitations and Future Directions

AICON models only the purely reactive component; an explicit dual-process model integrating both AICON’s gradients and state-based planning remains for future research. The data is also limited to a single, standardized ToL problem set and convenience clinical samples. Further work will be necessary to validate these results across different task topologies, environments, and patient populations. Finally, direct neural measurements (e.g., imaging, lesion mapping) could provide converging evidence for the alignment between AICON’s architecture and the biological substrates of planning and reactive action in the human brain.

Conclusion

This research demonstrates that a robotics-inspired, reactive control architecture predicts the specific problem difficulty ordering observed in planning-impaired populations performing the Tower of London task, outperforming both structural and algorithmic planning baselines in these groups. For healthy individuals, planning baselines retain their superiority, confirming the role of deliberative lookahead. The strong, a priori alignment between AICON’s structurally dictated failure modes and observed human deficits points to a genuine mechanistic correspondence, highlighting the value of reactive architectures in modeling components of human cognition that persist even as planning resources are compromised.

Reference: "No Plan, Yet Human: A Reactive Robotics Model Predicts Human Planning Failures on a Clinical Task" (2605.16514)