Conference on Robot Learning, CoRL 2026

SAFER-Nav: Enhancing Safety for Visual Robot Navigation via Segmentation-Aware Fine-Tuning

Anonymous Authors

Under review

Abstract

Recently, vision-based navigation models, particularly foundation models, have demonstrated promising performance by generating viable trajectories from RGB observations alone. However, even state-of-the-art (SOTA) transformer- and diffusion-based policies struggle to generalize in unfamiliar deployment environments containing unseen obstacles or shifted conditions. As a result, generated trajectories often remain goal-directed but unsafe. Existing efforts improve safety through external trajectory correction or internal geometric priors, yet the resulting policies are not trained to explicitly represent obstacle boundaries or traversable free-space structure. To address this limitation, we propose SAFER-Nav (Segmentation-Aware Fine-tuning for Enhancing Safety of Robot Navigation), a segmentation-aware navigation model that incorporates obstacle and free-space structure directly into the policy via fine-tuning and is designed to be compatible with diverse RGB-based navigation backbones. We evaluate SAFER-Nav against SOTA models, including ViNT, NoMaD, and their CARE-augmented variants, across multiple robot platforms, diverse indoor environments, and static and dynamic obstacle scenarios, demonstrating that SAFER-Nav reduces collision frequency while maintaining goal-reaching performance.

Video Comparisons

Static Obstacle Environment — LoCoBot

Static Obstacle Environment

Static obstacles evaluated on the LoCoBot platform.

SAFER-Nav (Ours)
ViNT
NoMaD
ViNT + CARE
NoMaD + CARE

Static Obstacle Environment — RoboMaster

Static Obstacle Environment

Static obstacles evaluated on the RoboMaster platform.

SAFER-Nav (Ours)
ViNT
NoMaD
ViNT + CARE
NoMaD + CARE

Static Obstacle Environment — TurtleBot4

Static Obstacle Environment

Static obstacles evaluated on the TurtleBot4 platform.

SAFER-Nav (Ours)
ViNT
NoMaD
ViNT + CARE
NoMaD + CARE

Dynamic Obstacle Environment — Corner-Appear

Dynamic Obstacle Environment

A dynamic obstacle appears from a corner along the robot's planned path.

SAFER-Nav (Ours)
ViNT
NoMaD
ViNT + CARE
NoMaD + CARE

Dynamic Obstacle Environment — Front-Approach

Dynamic Obstacle Environment

A dynamic obstacle approaches head-on toward the robot.

SAFER-Nav (Ours)
ViNT
NoMaD
ViNT + CARE
NoMaD + CARE

Static Obstacle Navigation Results

Comparison of navigation performance between baseline vision-based models and SAFER-Nav across three robot platforms and two different indoor environments. Metrics include goal arrival rate (%), collision counts per run, traveled distance (m), and completion time (s).

Robot & Model Environment TE Environment CY
Goal% ↑ #Coll. ↓ Dist. (m) Time (s) Goal% ↑ #Coll. ↓ Dist. (m) Time (s)
RoboMaster
SAFER-Nav (Ours) 0.95014.92 ± 0.6077.31 ± 6.38 0.90.1125.07 ± 0.49129.77 ± 4.36
ViNT 0.63.1714.59 ± 0.6779.99 ± 6.95 0.40.7526.20 ± 0.65128.63 ± 3.12
NoMaD 0.33.6716.38 ± 0.2087.50 ± 5.72 0.53.425.48 ± 0.78122.27 ± 3.70
ViNT + CARE 0.51.415.17 ± 0.6282.98 ± 4.22 0.1126.35138.25
NoMaD + CARE 0.2417.07 ± 1.0991.75 ± 3.18 0.1225.97131.19
TurtleBot4
SAFER-Nav (Ours) 0.850.3514.31 ± 0.9585.83 ± 7.31 0.90.2225.20 ± 0.80137.54 ± 5.31
ViNT 0.3315.21 ± 0.5284.03 ± 2.28 0.4126.12 ± 0.73141.89 ± 6.90
NoMaD 0.1723.42120.50 0.24.526.33 ± 0.15125.73 ± 5.49
ViNT + CARE 0.43.2514.35 ± 0.2885.75 ± 4.60 0.33.3326.84 ± 0.28134.37 ± 3.16
NoMaD + CARE 0.24.517.28 ± 2.31107.75 ± 7.07 0.21.526.08 ± 0.17147.95 ± 6.16
LoCoBot
SAFER-Nav (Ours) 0.950.2614.92 ± 0.2786.08 ± 3.35 0.950.1125.28 ± 0.87133.13 ± 3.26
ViNT 0.22.517.18 ± 0.2182.19 ± 2.74 0.60.1725.88 ± 0.82126.31 ± 6.09
NoMaD 0.22.518.78 ± 0.22102.34 ± 1.23 0.44.7526.14 ± 0.27129.94 ± 5.83
ViNT + CARE 0N/AN/AN/A 0.41.2525.29 ± 0.14142.72 ± 4.70
NoMaD + CARE 0N/AN/AN/A 0.3325.92 ± 0.49152.59 ± 2.18

Goal Arrival vs. Collision (Static Obstacle)

Each marker represents one method. The x-axis denotes average collision counts per run and the y-axis denotes goal arrival rate (%). Methods closer to the top-left corner indicate better overall performance.

Dynamic Obstacle Navigation Results

Number of trials where collisions occur (out of 10 trials) for each dynamic obstacle scenario. The dynamic obstacle is a teleoperated TurtleBot4 mobile robot.

Model (i) Corner-Appear (ii) Front-Approach
SAFER-Nav (Ours) 0/10 0/10
ViNT 5/10 9/10
NoMaD 6/10 8/10
ViNT + CARE 3/10 2/10
NoMaD + CARE 2/10 2/10