Lower sensing cost
Operates on egocentric RGB-D observations rather than full panoramic input.
ICRA 2026
Navigate beyond
what the robot sees.
DreamNav is a trajectory-based imaginative framework for zero-shot vision-and-language navigation using only low-cost egocentric observations.
1 The Hong Kong University of Science and Technology (Guangzhou) 2 Zhejiang Normal University * Equal contribution
Point-level
Trajectory-level
Long-horizon reasoning with aligned semantics
32.79Success Rate
28.95SPL
12 / 20Real-world success
RGB-DEgocentric only
01 / Motivation
From passive perception
to active imagination.
Existing zero-shot VLN systems often depend on expensive panoramic sensing and select isolated waypoints. This makes decisions short-sighted and can disconnect visual semantics from executable actions.
DreamNav instead predicts complete candidate trajectories, imagines their likely outcomes, and chooses actions with a longer planning horizon—all from a compact egocentric field of view.
Long-horizon planning
Reasons over candidate trajectories instead of one waypoint at a time.
Aligned decisions
Evaluates imagined outcomes while preserving instruction-action consistency.
02 / Framework
A four-part navigation loop.
Each module has a clear role, from correcting the current view to safely executing the best trajectory.
Module 01
EgoView Corrector
A Macro-Adjust Expert resolves large initialization errors, while a Micro-Adjust Controller corrects orientation drift after each action.
03 / EgoView Correction
Stay aligned at every step.
Egocentric agents can begin with the wrong heading or drift after executing an action. DreamNav handles both failure modes explicitly to keep instruction-relevant landmarks within view.
180°
Macro-Adjust Expert handles severe initial misorientation.
60°
Micro-Adjust Controller recovers from post-action deviation.
04 / Results
Strong without the panorama.
DreamNav improves navigation efficiency and success using only egocentric observations.
Success Rate
SPL
12 successes / 20 trials
SPL vs. prior egocentric methods
Real-world evaluation
One policy, four unseen spaces.
05 / Demo
DreamNav in motion.
Watch the complete project presentation and real-world navigation demonstrations.
06 / Poster
One-page project summary.
Open the ICRA poster for the full method diagram, ablations, simulation results, and real-world evaluation.
07 / Citation
Build on DreamNav.
If this work supports your research, please cite the paper.
@inproceedings{wang2026dreamnav,
title = {DreamNav: A Trajectory-Based Imaginative Framework
for Zero-Shot Vision-and-Language Navigation},
author = {Wang, Yunheng and Fang, Yuetong and Wang, Taowen and
Feng, Yixiao and Tan, Yawen and Zhang, Shuning and
Liu, Peiran and Ji, Yiding and Xu, Renjing},
booktitle = {2026 IEEE International Conference on Robotics
and Automation (ICRA)},
year = {2026}
}