Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds
Mingjian Gao1,2,* · Zhaocheng Li1,2,* · Haoyang Huang1,3,* · Wenqiao Zhang2,‡ · Yingjie Niu1,4,‡,† · Hao Zhou2 · Chao Li1 · Juncheng Li2 · Siliang Tang2 · Yueting Zhuang2
1Deeprobotics · 2Zhejiang University · 3University of California, San Diego · 4The Chinese University of Hong Kong
* Equal contribution · ‡ Corresponding authors · † Project lead
Overview · Key idea · EvoWorld-Bench · Results · Getting started
Robots often arrive in a world that no longer matches their last observation. A cup can be moved, a door can close, or an object can disappear while the agent is travelling. EvolvingNav treats this gap as a first-class part of navigation: it predicts what the world is likely to look like at arrival time, gathers evidence only when it is informative, and revises its plan when the evidence contradicts memory.
We introduce EvoWorld-Bench, a benchmark for persistent navigation in evolving 3D environments, and an event-driven agent that closes the loop between temporal memory, prediction, visibility-aware inspection, and replanning.
Most navigation systems treat an observation as a static fact. EvolvingNav separates three questions that are easy to conflate:
- Persistence: is the last-seen state still likely to hold?
- Relocation: if it changed, where could the object have moved under the observed routine?
- Evidence: does a new view actually rule out a hypothesis, or was the object simply not visible?
This separation lets the agent use a stale memory as a calibrated prior rather than as a hard-coded map. The largest gains occur when the environmental change has learnable regularity; under uncertainty, the agent keeps multiple hypotheses and uses route observations to resolve them.
| 1 · Remember Build causal entity versions from timestamped RGB-D histories. |
2 · Predict Forecast persistence and relocation at candidate arrival times. |
3 · Inspect Use visibility and depth coverage to qualify new evidence. |
4 · Replan Update belief, memory, and the next action after informative views. |
The method combines a continuous-time history encoder, a persistence–relocation belief, an arrival-time filter, and a frozen zero-shot vision-language controller. The controller is invoked inside an event-driven loop: action, visibility-qualified observation, belief update, and route revision.
EvoWorld-Bench turns human activity traces into executable evolving worlds. It preserves temporal histories, causal observability, controlled mobility regimes, and held-out transfer settings so that an agent cannot solve the task by treating the last observation as permanently true.
|
Figure 4 · Benchmark construction 
Human traces are normalized into a shared event schema, grounded in scenes, and instantiated as executable worlds. |
Figure 3 · Evaluation protocols 
N1–N5 and EQA probe prediction, search, replanning, online dynamics, transfer, and embodied question answering. |
| Dimension | Coverage |
|---|---|
| Scenes | 54 evolving indoor scenes |
| Tasks | 803,680 executable instances |
| World regimes | Static, routine, and random evolution with controlled mobility |
| Protocols | N1–N5 predictive navigation, search, replanning, online dynamics, and held-out transfer |
| Additional task | Embodied question answering (EQA) over evolving scenes |
The benchmark workflow is implemented in code/src/evoworld/ and includes generation, native Habitat RGB-D histories, public/private audits, evaluation adapters, and release-layout export.
On EvoWorld-Bench, EvolvingNav improves both the first destination chosen from stale memory and recovery within a search budget. Values below are reported as mean ± standard deviation across five seeds.
| Metric | EvolvingNav |
|---|---|
| First-Inspection Success Rate ↑ | 61.32 ± 3.24 |
| Search Success Rate ↑ | 86.18 ± 2.07 |
| SPL ↑ | 70.15 ± 1.74 |
Qualitative behavior: arrival-time prediction proposes plausible destinations, while informative observations suppress stale hypotheses and trigger recovery.
The ablation isolates the contribution of predictive belief, transition modeling, visibility-aware evidence, and event-driven replanning. Removing these components weakens either the initial inspection decision or the ability to recover after a world change.
We evaluate the same persistent-navigation loop on 64 matched LYNX M20 search episodes. Across indoor and outdoor scenes, the robot reaches 34.4% First-Inspection SR, 48.4% Search SR, and 24.3% Recovery SR, with 43.8 m mean travel. LYNX M20 is the primary quantitative platform; LYNX X30 and Lite3 provide matched transfer subsets with the same task protocol and perception interface.
Figure 14 · Indoor cup search
Predictive belief reaches the cup after it moves, while the last-seen policy revisits a stale location.
Figure 10 · Outdoor car search
The robot rejects the previous parking spot and searches for the car at its predicted current location.
The paper also reports two longer-horizon executions that test system-level orchestration beyond the quantitative predictive-search protocol.
Figure 15 · Indoor document delivery. The robot navigates to an office, verifies a person, collects documents with human assistance, and delivers them to a meeting room.
Figure 16 · Outdoor shared-bike search. The robot detects a blocked preferred route, replans through a detour, and finds the shared bikes.
The project page contains four HSSD videos plus HM3D and Habitat-GS demonstrations, an interactive HSSD belief explorer, benchmark figures, and the full paper narrative. An interactive lightweight demo is also available on Hugging Face.
.
├── index.html # Research project page
├── figures/ # Paper figures and web-ready previews
├── assets/ # Interactive explorer assets
├── main.pdf # Repository copy of the manuscript
└── code/
├── evolvingnav/ # Agent, memory, filter, controller, and Habitat runner
├── src/evoworld/ # EvoWorld-Bench generation and evaluation workflow
├── scripts/ # Dataset, training, calibration, and native rollouts
├── configs/ # Perception and benchmark configurations
└── tests/ # Agent and benchmark contract tests
The runnable implementation lives in code/. It targets Python 3.11 and Habitat-Sim 0.3.3.
git clone https://github.com/ZJU4EmbodiedAI/EvolvingNav.git
cd EvolvingNav/code
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
export PYTHONPATH=.:src:scripts
python -m pytest tests -qFor dataset paths, RGB-D assets, model checkpoints, calibration, training, and N1–N5 execution commands, see the detailed code README. The source tree keeps generated datasets, run logs, checkpoints, and credentials outside version control.
If EvolvingNav is useful for your research, please cite:
@article{gao2026evolvingnav,
title = {Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds},
author = {Gao, Mingjian and Li, Zhaocheng and Huang, Haoyang and Zhang, Wenqiao and Niu, Yingjie and Zhou, Hao and Li, Chao and Li, Juncheng and Tang, Siliang and Zhuang, Yueting},
journal = {arXiv preprint arXiv:2609.39166},
year = {2026}
}- Paper: arXiv:2609.39166
- Project page: zju4embodiedai.github.io/EvolvingNav
- Code: github.com/ZJU4EmbodiedAI/EvolvingNav/tree/main/code
- Interactive demo: Hugging Face Space