CanonNav: Disentangling Navigation Behavior from Camera Geometry in Cross-Platform Visual Navigation

Kim, Dong-Wook; Hwang, Ji-Hoon; Son, E-In; Oh, Mintaek; Seo, Seung-Woo

CanonNav: Disentangling Navigation Behavior from Camera Geometry in Cross-Platform Visual Navigation

Dong-Wook Kim, Ji-Hoon Hwang, E-In Son, Mintaek Oh, Seung-Woo Seo

Seoul National University, Republic of Korea

Paper Code arXiv

Motivation: (a) cross-platform demonstrations where the same metric trajectory projects to different image regions and different trajectories look alike in the image; (b) a robot whose imitation-only trajectory collides with a wall, while identifying a local-progress region and generating a safe path reaches toward the goal. — **(a)** On robots with different cameras, the same metric trajectory falls on different image regions, while different trajectories can look alike in the image. A policy trained directly on image-trajectory pairs must therefore infer camera geometry from every observation while it learns to navigate. **(b)** When the goal cannot be reached directly, the robot has to decide where to advance locally and how to get there safely. Demonstrations show only the motion the expert chose, so imitation alone leaves these decisions implicit and can steer straight into an obstacle.

Abstract

While visual navigation has advanced through imitation learning from cross-platform demonstrations, fully leveraging such data remains challenging. First, directly learning from image-trajectory pairs entangles navigation behavior with platform-dependent camera geometry. This hinders consistent learning by forcing the policy to implicitly infer camera geometry from visual observations, an inherently ill-posed problem. Second, imitation learning from demonstrated trajectories captures the expert's chosen motion but leaves the intermediate decisions underlying that motion implicit. To address these issues, we propose CanonNav, a visual navigation framework that disentangles navigation behavior from camera geometry and incorporates complementary planning supervision into learning from cross-platform demonstrations. CanonNav introduces camera geometry canonicalization, which transforms visual observations and trajectories into a camera-consistent representation space. Building on this representation, we derive safety and local-progress supervision using pseudo-labels from an offline traversability estimator. Safety supervision penalizes unsafe trajectories, while local-progress supervision guides where the robot should advance. Experiments across diverse camera configurations and environments show that, despite using only RGB at inference, CanonNav consistently outperforms RGB-based baselines and even surpasses RGB-D-based methods in challenging scenarios.

Method

CanonNav overview: (a) camera geometry canonicalization, (b) policy inference with diffusion trajectory, waypoint traversability and Scope-of-Reach heads, (c) navigation outputs, (d) training-time planning supervision from traversability estimation, a BEV collision map and its signed distance field. — **(a)** Images are warped to a canonical camera, and trajectories and goals are normalized by the camera height. **(b)** Image and goal features are fused by cross-attention and passed to the prediction heads. **(c)** The network predicts candidate trajectories, a traversability score for each waypoint, and the Scope-of-Reach (SoR); the robot follows the best path among the candidates. **(d)** During training, pseudo-labels from an offline traversability estimator yield a BEV collision map and signed distance field that supervise safety and the SoR target.

Results

Simulation Experiments

Simulation comparison of CanonNav with ViNT, NoMaD, OmniVLA-edge, LiMo, ViPlanner and NavDP in outdoor street and hospital scenes: predicted paths toward a red goal point. — Qualitative comparison in simulation across camera configurations (red dot: goal, green: selected path, orange: candidate paths, yellow: predicted SoR). CanonNav plans around obstacles toward the SoR, while baselines often head straight at the goal through cars, walls and people.

Method	Input	CitySim (Outdoor)						AWS Hospital (Indoor)
		6 m		12 m		20 m		6 m		9 m		12 m
		SR ↑	Coll. ↓	SR ↑	Coll. ↓	SR ↑	Coll. ↓	SR ↑	Coll. ↓	SR ↑	Coll. ↓	SR ↑	Coll. ↓
ViNT	RGB	93.8	0.69	63.0	4.27	18.6	8.03	62.1	9.23	38.8	13.07	30.0	13.10
NoMaD	RGB	98.0	0.43	61.6	4.29	24.5	7.00	65.6	8.15	40.2	12.22	29.8	13.30
OmniVLA-edge	RGB	89.4	0.76	61.4	2.87	28.5	4.26	72.2	6.11	41.7	10.98	31.1	11.68
LiMo	RGB	99.1	0.26	79.6	2.16	48.3	4.25	81.0	4.17	54.3	8.03	39.7	9.92
ViPlanner	RGB-D	81.6	0.61	62.9	3.00	39.3	4.87	86.6	2.09	69.3	4.23	52.2	6.04
NavDP	RGB-D	99.7	0.07	93.7	0.54	71.2	2.21	98.3	0.53	93.1	1.22	82.9	1.91
CanonNav (Ours)	RGB	99.4	0.20	94.1	0.53	86.4	0.89	97.1	1.26	92.2	1.77	85.5	2.37

Simulation results averaged over 18 camera configurations, for each subgoal interval. SR: subgoal success rate (%); Coll.: collisions per 100 m traveled. Best results in bold, second-best underlined. From Table 1 of the paper.

Real-world Experiments

BibTeX

@misc{kim2026canonnav,
      title={CanonNav: Disentangling Navigation Behavior from Camera Geometry in Cross-Platform Visual Navigation},
      author={Dong-Wook Kim and Ji-Hoon Hwang and E-In Son and Mintaek Oh and Seung-Woo Seo},
      year={2026},
      eprint={2608.30242},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2608.30242},
}