No Plan, Yet Human: A Reactive Robotics Model Predicts Human Planning Failures on a Clinical Task
Authors: Michael Migacev, Vito Mengers, Antonia Köngeter, Oliver Brock
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.
Large language models are increasingly used as planners for robotic systems, yet how safely they plan remains an open question. To evaluate safe planning systematically, we introduce DESPITE, a benchmark of 12,279 tasks spanning physical and normative dangers with fully deterministic validation. Across 23 models, even near-perfect planning ability does not ensure safety: the best-planning model fails to produce a valid plan on only 0.4% of tasks but produces dangerous plans on 28.3%. Among 18 open-source models from 3B to 671B parameters, planning ability improves substantially with scale (0.4-99.3%) while safety awareness remains relatively flat (38-57%). We identify a multiplicative relationship between these two capacities, showing that larger models complete more tasks safely primarily through improved planning, not through better danger avoidance. Three proprietary reasoning models reach notably higher safety awareness (71-81%), while non-reasoning proprietary models and open-source reasoning models remain below 57%. As planning ability approaches saturation for frontier models, improving safety awareness becomes a central challenge for deploying language-model planners in robotic systems.
Reasoning-based robotic policies using large language and vision-language models achieve strong semantic planning capabilities but mostly suffer from a high inference latency that limits practical real-time deployment. In this work, we observe that robotic reasoning workloads contain substantial temporal redundancy, where consecutive observations frequently produce identical actions and subgoals. Based on this insight, we present REIS, a human cognition inspired robotic decision-making framework that minimizes unnecessary reasoning while preserving semantic adaptability. REIS combines lightweight scene gating, KV-steered affordance routing, and deliberative reasoning to accelerate robotic control under embodied constraints. Experiments on ALFRED, and real-world robotic tasks demonstrate that REIS significantly suppresses reasoning overhead while maintaining competitive task performance.
When LLMs exhibit uneven performance across planning tasks, these gaps are often attributed to task difficulty. We argue that this explanation is incomplete, as task-level variation may reflect distinct latent planning competencies rather than differences along a single ability spectrum. We study this question on ACPBench-Hard by evaluating multiple LLM families under varying test-time reasoning budgets and applying a multidimensional item response theory model to uncover the latent competency structure underlying LLM planning. The analysis reveals two principal dimensions that shape planning performance: operational reasoning, the ability to evaluate local action applicability and immediate state transitions, and structural enumeration, the ability to reason about goal reachability and landmark structure. Operational reasoning improving under model scaling and longer reasoning traces, while structural enumeration remains comparatively insensitive. Our findings motivate competency-level evaluation of LLM planning, shifting the focus from whether models improve overall to which planning competencies improve, under what conditions, and why.