Physical Experience Builds Continual Spatial World Models

Yuhang Sun

cs.RONov 2, 2026

Physical Experience Builds Continual Spatial World Models

Authors: Yuhang Sun

Organizations: University of Washington

Abstract

World models are increasingly used to support prediction and planning in embodied agents, but the knowledge is usually established before deployment and then carried into later encounters. Continual operation faces a different problem. A familiar environment may change, and the agent must use previous experience without allowing old expectations to override current physical evidence. This study asks what repeated physical experience must contribute for a spatial world model to become progressively more useful rather than simply more persistent. A physics-grounded dynamic maze separates recurring spatial change from the continuous consequences of action. Cross-episode Spatial Change Memory learns relations among environmental transformations, while state-action-outcome experience refines predictions of motion. Structured experience progressively reduces the need for future route correction and transfers to spatial transitions not encountered during learning. The advantage disappears when past changes contain no information about future ones. Continuous navigation further shows that predictive memory changes behavior when a previous decision becomes invalid but leaves behavior unchanged when the decision remains correct. Physical tests expose a complementary limit: correct topology does not guarantee correct action prediction, and accumulated physical experience reduces this mismatch within the modeled dynamics. The results characterize continual world knowledge as experience that improves future prediction and action beyond stored episodes while remaining revisable by the physical outcomes it is intended to predict.

Figures & tables

Figure 1: Conceptual organization of the continual spatial world-model experiment. (a) Repeated navigation exposes the agent to related spatial changes across episodes. (b) Reactive updating observes a change, revises the current map, and recomputes the route, so each event is resolved from current evidence. (c) Spatial Change Memory pools transition observations across episodes to learn a relation shared across spatial locations. (d) The falsifiable boundary: structured change should support memory-based prediction, whereas random change ( $\varepsilon=0.5$ ) should not.

Figure 2: Reactive updating and accumulation across encounters. (a) Local adaptation accelerates as changed gates become observable sooner. (b) Shared SCM approaches the Bayes limit with accumulated episodes, while position-specific memory learns more slowly. (c) The same experience reduces later route corrections; error bars show SEM.

Figure 3: Transfer and the information boundary. (a) Shared SCM follows the theoretical prediction limit as environmental change becomes less predictable. (b) When past and future changes are independent, additional episodes do not improve prediction. (c) The shared relation transfers to a spatial transition excluded completely from training, unlike episodic and position-specific memories. Error bars show SEM across training replicates.

Figure 4: Predictive memory changes physical navigation before direct observation. (a) Direction-coded radial distance for representative matched seed 91000; positive and negative values denote clockwise and anticlockwise motion. The Reactive obsolete-pursuit interval is highlighted in red. Vertical markers show the structural change, SCM inference, and Reactive direct observation; the bracket marks the 17.73-s anticipatory lead. (b) Obsolete-target persistence for all 15 matched decision-changing episodes on a logarithmic axis. The distributions are completely separated: max(SCM) = 0.0083 s < min(Reactive) = 2.79 s; the 0.5-s line is a visual reference only. (c) Completion-time gain (Reactive $-$ SCM). Points are individual held-out episodes; the bar and whiskers show the mean and bootstrap 95% confidence interval for decision-changing episodes. Mean gain is 7.56 s, while all 15 decision-preserving episodes remain exactly at zero.

Test	Main comparison	Observation	Claim supported
Structured accumulation	0 vs 64 past episodes	accuracy 0.564 $\rightarrow$ 0.898; corrections 1.744 $\rightarrow$ 0.406	experience lowers future correction
Held-out transfer	Shared SCM vs episodic/position/reactive	0.898 vs approximately 0.566	transferable relation, not route cache
Zero-information control	I =0 with increasing history	accuracy remains approximately 0.50	memory cannot create predictive structure
Continuous navigation	memory vs reactive, decision-changing	obsolete-target time 7.806 s $\rightarrow$ 0.0083 s	prediction changes action before direct observation
Physical accumulation	inaccurate vs continually updated dynamics	parameter and prediction error decrease	action outcomes revise physical prediction

Table 1: Main evidence supporting the continual-world-model interpretation.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Quantity	Result	Role in the study
Effective mass check	expected 0.056 kg; measured 0.056 kg	verifies rolling-sphere reduction
Observation-radius effect	adaptation latency decreases monotonically with radius	establishes information-access limit
Distance-latency association	Spearman $\rho=0.812$ , $p=4.98\times 10^{-8}$	links reactive delay to information arrival
Structured SCM at 64 episodes	accuracy 0.898; Bayes limit 0.900	shows near-limit use of repeated structure
Random change at 64 episodes	accuracy approximately 0.50	zero-information negative control
Held-out shared SCM	accuracy 0.898 vs approximately 0.566 baselines	tests transfer beyond stored location

Appendix

Table 2: Selected quantitative checks and boundary conditions.

World Model Learning

World model learning builds predictive models of how environments change under actions to support planning, simulation, and decision-making.

See all 434 papers

Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds
Sep 30, 2026Mingjian Gao, Zhaocheng Li, Haoyang Huang +7Belief-Space Planning Embodied Navigation
Deeprobotics · Zhejiang University · University of California, San Diego +1
Neural Fields as World Models
Feb 21, 2026Joshua NunleyWorld Model Learning Action-Conditioned World Models
Luddy School of Informatics, Computing, and Engineering, Indiana University, Bloomington · Cognitive Science Program, Indiana University, Bloomington
From World Models to World Action Models: A Concise Tutorial for Robotics
Jul 1, 2026Xiaoxiong Zhang, Xiong Zeng, Wei ZhangWorld Model Learning World Models for Robotics
School of Automation and Intelligent Manufacturing, Southern University of Science and Technology, Shenzhen, China · LimX Dynamics, Shenzhen, China

Sep 30, 2026J/K move · Enter open · S save

Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds

Mingjian Gao, Zhaocheng Li, Haoyang Huang, Wenqiao Zhang+6

Deeprobotics · Zhejiang University · University of California, San Diego+1

Belief-Space Planning Embodied Navigation Robot Navigation

Open paper AskPDF arXiv

Persistent spatial memory enables embodied agents to navigate familiar environments across repeated visits. However, targets may move while unobserved, including during navigation, making remembered locations unreliable by the time an agent arrives. Despite advances in memory retrieval and state prediction, accounting for continued hidden world evolution and revising beliefs under limited visibility remain challenging. We study Evolving-World Navigation, where agents infer target locations from intermittent observations, predict their states at inspection time, and revise beliefs using visual evidence. We propose EvolvingNav, which constructs a time-indexed belief from timestamped 3D object histories through a structured persistence-relocation model. The belief distinguishes persistence at the last observed location from relocation to alternative locations and retains probability mass outside the known candidate set. An event-driven filter propagates the current belief as time elapses, forecasts target occupancy at candidate inspection times, and incorporates new RGB-D evidence. Negative observations downweight location hypotheses according to calibrated, visibility-conditioned detection probabilities, while evidence tracking prevents repeated use of the same observations. A frozen, zero-shot vision-language controller uses the updated belief to choose actions and replan. We further introduce EvoWorld-Bench, a benchmark grounded in human activity traces, comprising 54 scenes and 803,680 tasks with controlled changes before and during navigation. In simulation and real-robot experiments, EvolvingNav improves navigation success and search efficiency over the evaluated baselines. Paired experiments show the clearest gains under learnable temporal patterns, while ablations demonstrate the value of preserving uncertainty and incorporating visibility-aware evidence.

Abstract

World Model Learning

Explore similar work

Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds

Neural Fields as World Models

From World Models to World Action Models: A Concise Tutorial for Robotics