Self-motion as a structural prior for coherent and robust formation of cognitive maps

TL;DR

This work redefines self-motion as a structural prior that organizes the geometry and temporal coherence of learned cognitive maps, rather than merely updating position. It embeds a brain-inspired path-integration mechanism, inspired by CA3 circuitry, into a predictive-coding framework, yielding rapid convergence, robust local/topological preservation, and accurate global positioning under sensory ambiguity. Across diverse simulated and real-world environments, including a quadrupedal robot, the brain-inspired PI provides the strongest stability and generalization advantages, outperforming velocity-based and rate-based PI forms. The findings suggest a general principle for representation learning: trajectory-consistent internal dynamics can strengthen spatial intelligence in embodied agents and may extend to broader world-model and memory-augmented systems.

Abstract

Most computational accounts of cognitive maps assume that stability is achieved primarily through sensory anchoring, with self-motion contributing to incremental positional updates only. However, biological spatial representations often remain coherent even when sensory cues degrade or conflict, suggesting that self-motion may play a deeper organizational role. Here, we show that self-motion can act as a structural prior that actively organizes the geometry of learned cognitive maps. We embed a path-integration-based motion prior in a predictive-coding framework, implemented using a capacity-efficient, brain-inspired recurrent mechanism combining spiking dynamics, analog modulation and adaptive thresholds. Across highly aliased, dynamically changing and naturalistic environments, this structural prior consistently stabilizes map formation, improving local topological fidelity, global positional accuracy and next-step prediction under sensory ambiguity. Mechanistic analyses reveal that the motion prior itself encodes geometrically precise trajectories under tight constraints of internal states and generalizes zero-shot to unseen environments, outperforming simpler motion-based constraints. Finally, deployment on a quadrupedal robot demonstrates that motion-derived structural priors enhance online landmark-based navigation under real-world sensory variability. Together, these results reframe self-motion as an organizing scaffold for coherent spatial representations, showing how brain-inspired principles can systematically strengthen spatial intelligence in embodied artificial agents.

Figures (10)

• Figure 1: A brain-inspired path-integration prior for coherent cognitive-map formation.a Conceptual illustration of the hypothesized role of self-motion as a geometric and temporal scaffold that constrains the evolution of latent spatial representations when sensory cues are unreliable or ambiguous. b Self-motion is implemented as a path-integration (PI) mechanism and incorporated as a representation-level structural prior within the predictive-coding framework. c The brain-inspired PI neuron integrates recurrent spiking dynamics, analog membrane modulation, and adaptive firing thresholds to support stable, geometrically accurate, and capacity-efficient integration.
• Figure 1: Brain-inspired path integration converges rapidly and supports earlier emergence of structured spatial codes (256 hidden units, 3600 place cells).a Training dynamics show that the brain-inspired PI reaches a low-error regime by approximately 50k steps, whereas the rate-based PI requires nearly 500k steps to achieve comparable error levels. b At 500k training steps, the rate-based PI exhibits weak and spatially diffuse square-like periodicity. c By contrast, at 50k steps the brain-inspired PI already expresses clear hexagonal grid-like spatial patterns.
• Figure 2: Brain-inspired self-motion priors improve cognitive-map formation across diverse environments.a Three simulated environments used for evaluation: a highly aliased environment and a dynamic indoor environment designed in this work, and a naturalistic forest-cave-river benchmark gornet2024automated. b Self-motion priors consistently enhance local neighborhood structure, reflected in higher $\mathrm{H}(k)$ (F1-like measures of neighborhood preservation) and $\mathrm{LCMC}(k)$ (local continuity beyond random chance) scores. c Models trained with the self-motion prior encode more accurate global positional information, as assessed by a downstream predictor. d Next-image prediction error decreases across all environments, indicating stronger predictive continuity under the constraint.
• Figure 2: The brain-inspired path integration remains robust across latent-state integration factor $\alpha$ and exhibits a smooth transition in spatial response structure (256 hidden units, 3600 place cells).a Across a broad range of $\alpha$ values, the brain-inspired PI maintains consistently low integration error, indicating that performance is not dependent on precise hyperparameter tuning. b Spatial responses vary smoothly with $\alpha$, transitioning from pronounced grid-like periodicity toward increasingly localized, place-like patterns as $\alpha$ increases.
• Figure 3: Complementary CA3-derived properties shape an accurate and neurally consistent self-motion prior.a A controlled place-cell decoding task constructed from simulated rodent trajectories raudies2012modelingbanino2018vector. b Ground-truth place-cell activity and corresponding predictions produced by the PI module. c Test error and training dynamics for five PI variants, revealing substantial differences in convergence speed and integration accuracy. d Spatial coding patterns learned by each variant, showing that only variants incorporating both spiking dynamics and adaptive thresholds form coherent grid-like structure. Abbreviations: LT, learnable thresholds; AM, analog modulation.