🤖 AI

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

This study demonstrates that AICON, a reactive robotic framework devoid of lookahead planning, accurately predicts human planning difficulties and failure patterns on the Tower of London test, particularly for individuals with reduced cognitive capacity, suggesting that biological planning under constraint shifts toward reactive modes similar to those modeled in robotics.

Original authors: Michael Migacev, Vito Mengers, Antonia Köngeter, Oliver Brock

Published 2026-08-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Michael Migacev, Vito Mengers, Antonia Köngeter, Oliver Brock

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Human beings are remarkably good at solving puzzles that require a sequence of steps. Whether we are rearranging furniture in a room or navigating a complex route through a city, we often seem to look ahead, imagining future states to decide what to do next. This ability to plan is a cornerstone of human intelligence, but it is also fragile. When the brain is injured or affected by conditions like Parkinson's disease, this capacity can shrink, leaving people stuck on problems they could previously solve. Understanding exactly how and why these failures happen is crucial for developing better ways to diagnose and treat these conditions. For decades, scientists have tried to build computer models that mimic human thinking, usually by programming them to simulate future possibilities just as a person might. However, a new study suggests that the most accurate way to understand human struggle might not be to build a better planner, but to build a system that does not plan at all.

Researchers at the Technical University of Berlin and the Freie Universität Berlin turned to a simple, well-known puzzle called the Tower of London test to investigate this. In this test, a person is presented with three pegs of different heights and a set of colored beads. The goal is to move the beads from a starting arrangement to a specific target arrangement using the fewest moves possible. The rules are strict: a bead can only be moved if it is on top of a stack, and a peg can only hold a certain number of beads based on its height. While the rules are simple, the mental effort required varies wildly from one puzzle to another. Some arrangements are easy to solve, while others trap even healthy people in long, confusing sequences of moves. People with Parkinson's disease, mild cognitive impairment, or those who have suffered a stroke often fail these tests in very specific, predictable ways, getting stuck on certain types of puzzles while solving others.

To understand these patterns, the team applied a model called AICON, which was originally designed to help robots move objects in the real world. Unlike traditional computer programs that simulate the future to find a solution, AICON does not look ahead. It reacts only to the immediate situation. Imagine a robot that can only see the object directly in front of it and the goal right now; it calculates the steepest path toward the goal based on what is currently visible and takes a step in that direction. It repeats this process, constantly adjusting its path based on the new state of the world, without ever simulating a sequence of moves five steps into the future. The researchers used this reactive system to solve the same 24 Tower of London puzzles that were given to human participants. They then compared the computer's performance to the actual results from 295 people, including healthy volunteers and patients with various neurological conditions.

The results revealed a striking and counterintuitive pattern. When the researchers looked at the healthy volunteers, the reactive computer model was not the best match. Instead, the healthy people's performance was better explained by models that did look ahead and plan, suggesting that healthy humans do indeed use forward thinking to solve these puzzles. However, the story changed completely when the researchers looked at the groups with reduced planning capacity. For people with Parkinson's disease, mild cognitive impairment, or a history of stroke, the reactive model was far more accurate than any planning-based model. The computer that did not plan predicted the specific difficulties these patients faced better than a computer that tried to plan. It correctly identified which puzzles would be hardest for them and, crucially, predicted the specific types of mistakes they would make.

This finding supports a long-standing hypothesis that was first noted by the creators of the AICON model before this study began. They had observed that the model's own weaknesses—specifically, its struggle with puzzles that require managing conflicting goals—mirrored the struggles of patients with Parkinson's disease. The new data confirms that as the brain's ability to plan diminishes, human behavior shifts toward a reactive mode, relying on immediate cues rather than future simulations. The study did not just show that the model could solve the puzzles; it showed that the model failed in the exact same way specific groups of people failed. This alignment suggests that the reactive strategy is not just a fallback for broken brains, but a fundamental part of how biological systems process information. It implies that even when we are healthy, we likely use a mix of planning and reacting, but when our planning capacity is compromised, the reactive side takes over, and the errors we make reveal the underlying structure of that system.

The researchers were careful to note that this does not mean planning is unimportant or that the reactive model is a perfect replica of the human mind. The model cannot account for learning, individual differences, or the deliberate planning strategies that healthy people use. It captures only one component of the puzzle-solving process. However, the fact that a system built for robotic manipulation, with no knowledge of human neuroscience, could predict the specific failure patterns of patients with brain disorders is a powerful indicator. It suggests that the core logic of how biological systems organize information is deeper and more universal than previously thought. By studying how a simple, reactive machine fails, scientists may have found a new window into how the human brain works when it is under stress, offering a potential path toward more precise diagnostic tools and targeted therapies for those who struggle with sequential planning.

Technical Summary: No Plan, Yet Human

Problem Statement
Understanding why specific sequential planning problems are harder than others requires models that capture fine-grained difficulty orderings rather than just average performance. While structural task parameters (e.g., minimum move count) explain aggregate performance, they fail to account for the specific cognitive processes generating difficulty. Furthermore, existing computational models of the Tower of London (ToL) test largely assume explicit lookahead planning. This assumption makes it difficult to resolve the tension between planning and reactive strategies, particularly in healthy populations where both may be intertwined. The challenge is to identify a model that can reproduce the specific failure patterns of humans with reduced planning capacity (e.g., Parkinson's disease, mild cognitive impairment, stroke) without relying on explicit forward search or prior knowledge of human cognition.

Methodology
The authors apply Active InterCONnect (AICON), a reactive gradient-descent framework originally developed for robotic manipulation, to the Tower of London test.

The Model: AICON encodes the ToL rules as a network of interconnected estimators (State, Movable, Free) and active interconnections. Instead of simulating future states, it uses gradient descent on a goal cost function to select actions based on the current world state. It computes gradients through paths representing direct moves or subgoal resolution (e.g., freeing a target field). The model selects the steepest available gradient at each step, implicitly identifying actionable subgoals without explicit lookahead.
Data: The model was evaluated against behavioral data from Köstering et al. [10,11], comprising 24 standardized ToL problems across four groups: Healthy Controls (HC), Parkinson's Disease (PD), Mild Cognitive Impairment (MCI), and Stroke patients.
Evaluation: The primary metric was Kendall's $\tau$ , measuring the rank correlation between the model's predicted difficulty ordering and human behavioral data (success rate and average additional moves).
Generalization: To prevent overfitting, a "leave-two-out" procedure was used. Parameters ( $\alpha, \beta$ ) were tuned on 22 problems, and the model's ability to order the remaining 2 held-out problems was tested. This was repeated across 120 random splits.
Baselines: AICON was compared against:
1. Structural Parameters: The optimal number of moves required.
2. Planning Baselines: Unidirectional and Bidirectional Breadth-First Search (BFS) algorithms, representing explicit planning strategies.

Key Results

Reproduction of Difficulty Ordering: AICON achieved consistent positive Kendall's $\tau$ across all four groups on both training and held-out test sets, outperforming structural parameters (optimal moves) in capturing fine-grained difficulty ordering.
Dissociation Between Groups:
- Healthy Controls: For healthy participants, the planning baseline (Bidirectional BFS) matched or exceeded AICON's performance. This aligns with the hypothesis that healthy individuals engage in meaningful lookahead planning.
- Planning-Impaired Groups: For groups with reduced planning capacity (PD, MCI, Stroke), AICON significantly outperformed the planning baselines, particularly for the PD group regarding "additional moves." The planning baseline failed to capture the difficulty ordering for these groups, whereas AICON maintained robust generalization.
Failure Modes: The model's failure profile—specifically its struggle with problems requiring the anticipation of conflicting subgoal hierarchies—mirrored the characteristic errors of Parkinson's patients. This alignment was strongest in the "hard vs. hard" problem splits, where gradient signals weaken for long subgoal chains.

Key Contributions

Application of AICON to Clinical Tasks: This is the first application of the AICON framework to the Tower of London test, demonstrating that a model built for robotic manipulation can capture human sequential planning behavior without being designed for it.
Mechanistic Validation of Reactive Strategies: The study provides evidence that human behavior shifts toward reactive strategies as planning capacity is reduced. The fact that a reactive model outperforms a planning model for impaired groups (while a planning model fits healthy controls better) suggests a theoretical dissociation rather than a modeling artifact.
Prospective Prediction Confirmation: The results confirm a prediction made in the original AICON paper [16]: that the model's specific failure modes (struggling with conflicting subgoals) resemble the planning deficits of Parkinson's patients. This was a prospective observation, not a post-hoc fit.
Superiority of Mechanistic over Descriptive Models: The study demonstrates that a mechanistic model (AICON) captures fine-grained difficulty orderings better than descriptive parameters (optimal moves), which fail to distinguish problems with identical structural parameters but different cognitive demands.

Significance and Claims
The paper claims that AICON captures a "real" aspect of how biological systems are organized, specifically the reactive component of sequential problem solving. The core argument is that as planning capacity diminishes (due to PD, stroke, or MCI), human behavior naturally shifts toward the reactive mode modeled by AICON.

The significance lies in the mechanistic alignment: because AICON was not tuned to human data or neuroscience, its ability to reproduce the specific failure profile of basal ganglia dysfunction suggests that the reactive gradient mechanism is a fundamental component of human cognition. This extends a broader pattern where AICON has successfully modeled biological behavior in visual perception, eye movements, and collective opinion dynamics.

The authors remain modest about the scope, noting that AICON models only the reactive component and does not yet account for deliberate planning, learning, or individual differences. They conclude that while the model is not a complete account of human planning, it successfully isolates and validates the reactive strategy that becomes dominant when planning resources are compromised.

Technical Summary: No Plan, Yet Human

More like this