bioRxiv Science⌕ Search
Home/Archive/bioRxiv

bioRxiv · 10.64898/2026.06.03.729790

How a Predictive State Observer Can Self-Adapt Its Sensory Prediction-Error Correction Gain: Closed-Loop Evidence from a Muscle-Driven Reaching Task

Kobayashi, J.
View original source ↗Export metadata ↓

Abstract

We ask how a forward-model-based predictive state observer should set its sensory prediction-error correction gain during muscle-driven reaching, and whether that gain can be adapted from agent-available signals -- innovation history and per-episode reaching outcome -- rather than from swept oracle labels. We evaluate a residual-MLP forward model in a 34-muscle MyoSuite arm on an IK-reachable below-shoulder task, deployed in closed loop with a stabilized endpoint probe controller that uses non-negative least-squares muscle routing and a virtual target ramp; the controller is a stabilized probe for evaluating state-estimation effects, not a biological motor planner. A swept fixed-gain closed-loop oracle reveals a delay-dependent correction structure: with no sensory delay, intermediate correction gains are best (K = 0.25-0.50), whereas with 18-step delay observation-heavy correction wins (K = 1.0). The forward-model-only K = 0 ablation is not the oracle: it is systematically worse than the best fixed K by 1.9-6.1 cm and shows large NNLS controller residuals caused by long-horizon autoregressive drift; we therefore report K = 0 as a diagnostic. Outcome-trained reliability-adaptive observers improve the delayed regime by 1.9-2.5 cm over default reliability while remaining neutral in no-delay cells, where the oracle is already intermediate. A feature-conditioned {beta} adapter that maps cell-level innovation statistics to per-field gain parameters nearly matches a per-cell trained diagnostic in 5/6 cells, but both remain 1.4-1.8 cm worse than the swept fixed-K oracle at 18-step delay. These results separate the delay-dependent correction structure, the forward-model-only failure mode of K = 0, and the remaining limits of agent-available adaptive correction.

Explore related subjects

neurosciencePreprint

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Kobayashi, J.. 2026-06-08. How a Predictive State Observer Can Self-Adapt Its Sensory Prediction-Error Correction Gain: Closed-Loop Evidence from a Muscle-Driven Reaching Task. https://doi.org/10.64898/2026.06.03.729790

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Same Movements, Different Skills: How Context Shapes Learning in Humans and Neural Networks

What determines whether experiences are integrated into shared representations or split? Motor skill learning offers a revealing test: when two tasks require an identical sequence of movements, both intuition and theory suggest that practice in one should generalize to the other. Yet we find that changing context can make identical movements different skills. In simple recurrent networks trained on the same sequence across contexts, splitting was greatest when strong contextual inputs coincided with a high initial weight scale; neither alone sufficed. Greater representational separation coincided with separate context-specific learning curves. Humans practicing the same finger sequence while alternating between two performance-irrelevant contexts likewise exhibited context-specific curves, rather than the single continuous trajectory observed when practicing in a single context. State-space modeling favored split learning states in humans and very weakly coupled states in the most context-dependent networks. These findings have broader implications for understanding how learning generalizes across contexts.

neuroscience↗
bioRxiv2026

Postpartum resource scarcity reshapes maternal care, offspring prioritization, and the steroid-sensitive medial preoptic area

Parental care requires balancing offspring needs with environmental constraints, yet how resource scarcity alters the organization of maternal responses remains poorly understood. Here, we examined maternal behavior, offspring prioritization, and steroid receptor-associated extracellular matrix organization in postpartum rats exposed to limited bedding and nesting (LBN) or control conditions with ample nesting materials. LBN increased pup-directed and reduced self-directed behaviors but also increased behavioral entropy, a measure of behavioral fragmentation, indicating less predictable caregiving sequences. LBN dams preferentially retrieved male pups first, whereas control dams preferentially retrieved females first. LBN also increased the proportion of pup-investigation time directed toward males compared with controls. In offspring on postnatal day 10, corticosterone was elevated in female but not male LBN pups, a pattern that paralleled the maternal male bias. To identify associated neurobiological changes, we assessed steroid receptors and extracellular matrix organization in the medial preoptic area (mPOA), a key region for maternal behavior. LBN increased numbers of estrogen receptor (ER)- and androgen receptor (AR)-immunoreactive cells, reduced perineuronal net (PNN) abundance, and decreased receptor-positive cell enwrapment. Because PNNs constrain structural plasticity, these findings suggest that the postpartum environment alters the organization of steroid receptor-associated extracellular matrix, potentially modulating experience-dependent plasticity in neurosteroid-sensitive circuits. In sum, LBN increases the levels of maternal engagement while reshaping its temporal organization and allocation, accompanied by selective remodeling of steroid-sensitive preoptic architecture. Therefore, maternal adaptation to adversity may be better understood as changes in caregiving allocation strategies rather than a simple gain or loss of maternal care.

neuroscience↗
bioRxiv2026

The Computational Grammar of Human Behaviour in a Biological Environment

Human behaviour unfolds as an extended stream of spontaneous actions shaped by environmental demands and the agent's goals. The organisation of these actions, situated between setting abstract goals and concrete movement execution, remains poorly characterised. Here, we introduce DynaClade, an unsupervised temporal action segmentation algorithm for human full-body motion capture data, and apply it to freely moving participants navigating ecologically realistic risky foraging environments in virtual reality. DynaClade identifies discrete actions whose kinematic characteristics and spatial distributions correspond to environmental affordances, supporting the validity of the approach. Among these actions, we discovered a hitherto undescribed movement pattern defined by increased body sway without locomotion. Action sequences were highly non-random and showed clear history dependence beyond first-order Markov structure, most of which was explained by a second-order prediction. This higher-order structure was well captured in a hidden Markov model, in which some of the observed actions were associated with more than one latent state. Each of these latent states corresponded to the same action in a distinct action context. These findings suggest that spontaneous human actions are predominantly selected based on the current action and context, rather than following pre-planned or habitually assembled sequences.

neuroscience↗