Organizations: Department of Electrical and Computer Engineering, Seoul National University (SNU) and Automation and Systems Research Institute (ASRI) and Sequor Robotics Inc., Seoul, Korea (Republic of). · Department of Electrical and Computer Engineering, Seoul National University (SNU) and Automation and Systems Research Institute (ASRI), Seoul, Korea (Republic of). · Interdisciplinary Program in Artificial Intelligence, Seoul National University (SNU) and Automation and Systems Research Institute (ASRI) and Sequor Robotics Inc., Seoul, Korea (Republic of). · Sequor Robotics Inc., Seoul, Korea (Republic of).
Safe visual navigation is critical for indoor mobile robots operating in cluttered environments. Existing benchmarks, however, often neglect collisions or are designed for outdoor scenarios, making them unsuitable for indoor visual navigation. To address this limitation, we introduce the reactive visual navigation benchmark (RVN-Bench), a collision-aware benchmark for indoor mobile robots. In RVN-Bench, an agent must reach sequential goal positions in previously unseen environments using only visual observations and no prior map, while avoiding collisions. Built on the Habitat 2.0 simulator and leveraging high-fidelity HM3D scenes, RVN-Bench provides large-scale, diverse indoor environments, defines a collision-aware navigation task and evaluation metrics, and offers tools for standardized training and benchmarking. RVN-Bench supports both online and offline learning by offering an environment for online reinforcement learning, a trajectory image dataset generator, and tools for producing negative trajectory image datasets that capture collision events. Evaluations demonstrate that policies trained on RVN-Bench generalize effectively across unseen simulated environments. Furthermore, initial physical experiments using a Jackal UGV indicate promising sim-to-real transfer. Code and additional materials are available at: https://sequor-robotics-research.github.io/projects/RVN-Bench/.
Figures & tables
Fig. 1: Overview of the RVN-Bench. The benchmark is designed for indoor mobile robots and focuses on collision-aware visual navigation, where an agent must reach sequential goal positions using only visual observations while avoiding collisions. Built upon Habitat 2.0 and utilizing HM3D scenes, it provides high-quality visual observations, an RL environment for training and evaluation, and pipelines for collecting trajectory image datasets, including negative (collision-inducing) trajectories.
Benchmark
Goal Type
Domain
Realistic Visual Rendering
Detect Collision
Lifelong Evaluation
Dynamic Obstacles
CARLA [ 5 ]
High-level command
Autonomous driving
✓
✓
✓
MetaUrban [ 13 ]
Point
Outdoor
✓
✓
Habitat Challenge [ 7 , 18 ]
Point, object, image
Indoor
✓
HM3D-OVON [ 19 ]
Language
Indoor
✓
GOAT-Bench [ 11 ]
Object, image, language
Indoor
✓
✓
HabiCrowd [ 10 ]
Point, object
Indoor
✓
▲
✓
TABLE I: Comparison of Visual Navigation Benchmarks.
Fig. 2: Trajectory image dataset generation process. We generate trajectory image datasets through the following steps. (a) Start and goal positions are randomly sampled in the scene. (b) The shortest path between these positions is computed using the ground truth occupancy map padded with a margin larger than the agent’s radius. (c) The agent follows this path using discrete actions. (d) If the agent reaches the goal position without collision, its positions, yaws, and image observations at each time step are recorded. For negative trajectory image dataset generation , the occupancy map is padded with a margin smaller than the agent radius to create path that is unsafe for the agent in step (b) . If the agent collides with an obstacle at t=tc , the history of its positions, yaws, and image observations from ti to tf is recorded, where ti=max(0,tc−kpre) , tf=min(T,tc+kpost) , and It is set to Itcfor t>tc .
Category
Method
Train
Validation
Test
SR 1 ↑
E(G)↑
CPK ↓
SR 1 ↑
E(G)↑
CPK ↓
SR 1 ↑
E(G)↑
CPK ↓
IL
ViNT-PointGoal [ 2 ]
0.097
0.11
442.3
0.090
0.11
468.3
0.093
0.10
465.4
NoMaD-PointGoal [ 3 ]
0.758
4.01
32.5
0.752
3.69
35.7
0.751
4.52
31.0
NoMaD-Neg
0.750
3.83
31.2
0.757
3.97
28.6
0.760
4.61
25.8
Safe-RL
PPO-Lagrangian [ 17 ]
0.826
7.88
15.7
0.803
7.61
16.9
0.805
9.02
13.7
RL
PPO [ 16 ]
0.824
8.70
15.6
0.808
8.54
16.4
0.819
8.68
15.5
TABLE II: RVN-Bench Results.
Fig. 3: Overview of the NoMaD-Neg framework. (a) Two NoMaD-PointGoal model, NoMaD e and NoMaD n , are trained separately with expert and negative trajectory datasets. (b) NoMaD e and NoMaD n predict expert trajectories ae,i∈aE and negative trajectories an,i∈aN , respectively. The expert trajectory ae∗ with minimal CoR(ae,i,aE,aN) is then selected.
Method
Observation Type
Train
Validation
Test
SR 1 ↑
E(G)↑
CPK ↓
SR 1 ↑
E(G)↑
CPK ↓
SR 1 ↑
E(G)↑
CPK ↓
DD-PPO [ 15 ]
RGB
0.899
13.28
9.2
0.865
11.86
10.5
0.886
13.90
8.7
RGB + Predicted Depth [ 25 ]
0.937
22.70
3.5
0.909
19.88
4.0
0.928
20.79
3.6
RGB + GT Depth
0.946
20.79
2.0
0.920
21.90
2.2
0.940
22.99
1.8
TABLE III: Ablation study on effect of observation types of DDPPO.
Fig. 4: Real-world evaluation of NoMaD-PointGoal with different training datasets. (a) Clearpath Jackal UGV platform [ 31 , 32 , 33 ] used in the experiments. (b) Illustration of evaluation results in a house environment. Shown are the executed trajectories of the NoMaD-PointGoal agent trained on the real-world dataset (pink), trained on the simulation dataset (yellow), and trained on the combined real-world and simulation datasets (blue).
Model
Dataset
SR 1 ↑
E(G)↑
CPK ↓
NoMaD-PointGoal [ 3 ]
Real
0.30
0.30
223.5
Sim
0.60
1.05
196.4
Real+Sim
0.75
1.30
191.4
TABLE IV: Real-world evaluation results of NoMaD-PointGoal models trained with different training datasets
Robots deployed in delivery, campus, and emergency-response settings often need to navigate from buildings to streets within a single continuous episode. Existing benchmarks usually evaluate indoor and outdoor navigation separately, and many abstract away robot execution, leaving exit finding, boundary traversal, adaptation, and kinodynamic failures underexplored. We introduce NavVerse, a physics-enabled benchmark for indoor-to-outdoor embodied navigation. NavVerse contains 100 indoor scenes, 50 urban outdoor scenes, and 50 indoor-to-outdoor scenes, and 10,000 episodes spanning Object Navigation, Vision-and-Language Navigation, and Place Navigation tasks, where agents search for semantic points of interest such as restaurants or banks. Agents are evaluated through executable robot interfaces using task-success, path-efficiency, and safety metrics. Zero-shot experiments with RL, VLA, and modular baselines show that current agents remain far from solving cross-context navigation: end-to-end VLAs obtain the highest zero-shot success, while the modular method provides the strongest safety profile. PlaceNav further reveals a clear drop from outdoor to indoor-to-outdoor scenes, indicating that adaptation remains major bottleneck.
Junzhe Wu, Yue Hu, Zeyu Han +4
University of Michigan, Ann Arbor · Tsinghua University
Visual Navigation Models (VNMs) promise generalizable, robot navigation by learning from large-scale visual demonstrations. Despite growing real-world deployment, existing evaluations rely almost exclusively on success rate, whether the robot reaches its goal, which conceals trajectory quality, collision behavior, and robustness to environmental change. We present a real-world evaluation of five state-of-the-art VNMs (GNM, ViNT, NoMaD, NaviBridger, and CrossFormer) across two robot platforms and five environments spanning indoor and outdoor settings. Beyond success rate, we combine path-based metrics with vision-based goal-recognition scores and assess robustness through controlled image perturbations (motion blur, sunflare). Our analysis uncovers three systematic limitations: (a) even architecturally sophisticated diffusion and transformer-based models exhibit frequent collisions, indicating limited geometric understanding; (b) models fail to discriminate between different locations that are perceptually similar, however some semantics differences are present, causing goal prediction errors in repetitive environments; and (c) performance degrades under distribution shift. We will publicly release our evaluation codebase and dataset to facilitate reproducible benchmarking of VNMs.
Visual navigation is a core capability for mobile robots, yet end-to-end learning-based methods often struggle with generalization and safety in unseen, cluttered, or narrow environments. These limitations are especially pronounced in dense indoor settings, where collisions are likely and end-to-end models frequently fail. To address this, we propose SaferPath, a hierarchical visual navigation framework that leverages learned guidance from existing end-to-end models and refines it through a safety-constrained optimization-control module. SaferPath transforms visual observations into a traversable-area map and refines guidance trajectories using Model Predictive Stein Variational Evolution Strategy (MP-SVES), efficiently generating safe trajectories in only a few iterations. The refined trajectories are tracked by an MPC controller, ensuring robust navigation in complex environments. Extensive experiments in scenarios with unseen obstacles, dense unstructured spaces, and narrow corridors demonstrate that SaferPath consistently improves success rates and reduces collisions, outperforming representative baselines such as ViNT and NoMaD, and enabling safe navigation in challenging real-world settings.
Lingjie Zhang, Zeyu Jiang, Changhao Chen
PEAK-Lab, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 511453, China