Driving · 2026-11-04
Physical Experience Builds Continual Spatial World Models
Yuhang Sun
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors study when repeated physical experience makes a spatial world model more useful for continual navigation rather than just more persistent. In a dynamic maze, cross-episode Spatial Change Memory learns transferrable relations between environmental changes, while state-action-outcome experience refines motion prediction. Structured predictability improves future routing and action selectively; advantages vanish when past changes are uninformative.
Why it matters
The paper proposes a functional criterion for continual world models: improvement should reduce future inference/correction and transfer to unseen situations, while remaining revisable by physical outcomes. It separates persistence from learning, tests an information-theoretic boundary for when memory helps, and evaluates action-level consequences with decision-changing vs decision-preserving episodes.
Method
- Use a physics-grounded dynamic maze where gates change between encounters; separate recurring spatial change from action consequences using cross-episode SCM and a state-action-outcome dynamics model.
- Train SCM to pool evidence about relations among environmental transformations, while updating action-conditioned motion predictions from accumulated state-action-outcome transitions.
- Evaluate continual value with structured vs random change (predictability sweep), a held-out spatial transition transfer test, matched continuous-navigation decision-changing/preserving episodes, and physical-model mismatch identification.
Limitation
The controlled design limits claims: spatial change is binary and structured; a visible gate is observed without recognition error; the physical state is low-dimensional; and parameter learning assumes the correct equation family. The results do not establish general visual place recognition, arbitrary dynamics discovery, or robustness to unrestricted distribution shift.
Abstract (from arXiv)
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.
Related papers
- PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation
- Action-Consequence Alignment for Reliable Planning and Self-Improving in Latent World Models
- MOST: Memory of Object States and Task Progress for Long-Horizon Mobile Manipulation
- AFTER: A World Model with Adaptive Fast--Slow Timescales for Error-Driven Retention
- EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-04 · Indexed by Pub-AI: 2026-11-04 UTC
Physical Experience Builds Continual Spatial World Models
First announced by arXiv: 2026-11-04 · Original record ↗
- arxiv
- 2611.04016
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-04 · First detected: 2026-11-04 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.04016v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
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.
