Semantic occupancy mapping supports embodied perception during exploration and detailed scene understanding afterward, giving rise to two complementary requirements: online incremental mapping for maintaining an up-to-date global map under causal constraints, and offline global map refinement for improving the completed scene representation with additional computation. We present VEOcc++, a unified voxel-centric framework that connects both stages through a persistent sparse voxel feature map. During exploration, the map expands with newly observed regions and integrates aligned local features through reliability-guided updates, while context-aware decoding produces the current semantic occupancy map. After exploration, cached local features are consolidated as complementary evidence and combined with the completed online representation for multi-scale spatial refinement. On EmbodiedOcc-ScanNet, VEOcc++ achieves 67.12% IoU and 58.40% mIoU, establishing new state-of-the-art performance for online incremental mapping. With the local predictor fixed, our online mapping pipeline outperforms VEOcc fusion by 3.70 IoU and 3.74 mIoU points, demonstrating substantial gains from the proposed global mapping design. Offline global map refinement further improves the completed map to 67.79% IoU and 59.37% mIoU using only the observations acquired during exploration. Zero-shot deployment on a mobile robot further demonstrates transferability to unseen environments, onboard feasibility, and scalability to large-scale indoor mapping. Project page: https://wryzju.github.io/VEOccpp/.
Figures & tables
Figure 1: Overview of the two-stage mapping setting in VEOcc++. Online incremental mapping sequentially integrates observations under causal constraints, while offline global map refinement exploits complementary local evidence and broader spatial context over the completed map. Zoomed examples highlight representative improvements after refinement.
Figure 2: Overall framework of VEOcc++. During online incremental mapping, local voxel features are aligned with the global map and accumulated in a Persistent Voxel Feature Map through reliability-guided updates, while Context-Aware Decoding produces the current occupancy map. During offline global map refinement, Local Evidence Aggregation consolidates cached local features as complementary evidence, which is combined with the completed online representation for Global Spatial Refinement.
Method
IoU
mIoU
■ ceiling
■ floor
■ wall
■ window
■ chair
■ bed
■ sofa
■ table
■ tvs
■ furniture
■ objects
EmbodiedOcc ( Wu et al., 2025 )
53.95
45.48
40.90
50.80
41.90
33.00
41.20
55.20
61.90
43.80
35.40
53.50
42.90
EmbodiedOcc++ ( Wang et al., 2025 )
54.90
46.20
36.40
53.10
41.80
34.40
42.90
57.30
64.10
45.20
34.80
54.20
44.10
RoboOcc ( Zhang et al., 2025 )
56.48
47.67
45.36
53.49
44.35
34.81
43.38
56.93
63.35
46.35
36.12
55.48
44.78
TGSFormer ( Qian et al., 2026 )
64.42
54.73
54.20
61.30
50.50
41.30
49.50
64.50
69.30
52.40
47.10
63.00
48.80
GPOcc (VGGT) ( Zhou et al., 2026a )
63.14
56.19
51.67
59.93
52.07
46.44
51.35
64.45
69.47
54.30
51.76
63.29
53.36
GPOcc++ (VGGT) ( Zhou et al., 2026b )
63.72
57.41
56.62
61.40
50.63
48.13
53.12
66.34
71.05
56.48
48.95
64.48
54.22
Table 1: Local Occupancy Prediction results of embodied occupancy mapping methods on Occ-ScanNet. The best results are highlighted in bold , and the second-best results are underlined .
Method
IoU
mIoU
■ ceiling
■ floor
■ wall
■ window
■ chair
■ bed
■ sofa
■ table
■ tvs
■ furniture
■ objects
EmbodiedOcc ( Wu et al., 2025 )
51.52
42.53
22.70
44.60
37.40
38.00
50.10
56.70
59.70
35.40
38.40
52.00
32.90
EmbodiedOcc++ ( Wang et al., 2025 )
52.20
43.60
27.90
43.90
38.70
40.60
49.00
57.90
59.20
36.80
37.80
53.50
34.10
RoboOcc ( Zhang et al., 2025 )
53.30
44.05
21.94
44.57
39.54
38.48
51.28
57.04
63.09
36.70
43.05
54.42
34.38
TGSFormer ( Qian et al., 2026 )
54.42
45.29
31.00
39.90
39.90
32.70
52.00
66.10
53.00
47.30
38.70
59.60
37.90
GPOcc (VGGT) ( Zhou et al., 2026a )
61.41
55.39
42.61
51.35
51.49
48.72
54.32
67.91
70.73
52.94
54.75
64.76
49.67
GPOcc++ (VGGT) ( Zhou et al., 2026b )
62.50
57.41
48.00
52.14
54.70
53.39
54.71
68.82
71.03
54.27
55.23
66.81
52.42
Table 2: Global Occupancy Mapping results on EmbodiedOcc-ScanNet. The best results are highlighted in bold , and the second-best results are underlined . † denotes VEOcc using the same local predictor as VEOcc++.
Figure 3: Progressive online incremental mapping and offline map refinement. Each row shows the map evolution from t=1 to t=T , followed by the refined result and ground truth. Red dashed boxes indicate the enlarged regions.
Figure 4: Qualitative comparison of online incremental mapping. Highlighted regions are enlarged for detailed comparison.
Figure 5: Zero-shot real-world mapping results. The upper and lower rows show the Reading Room and Laboratory Floor, respectively. Scene reconstructions are provided only for spatial reference and are not used by the occupancy predictor. The global maps demonstrate online incremental mapping at different scene scales, while offline refinement further improves the Reading Room result. A–C highlight representative local regions.
Setting
Online s/frame
Online s/scene
Offline s/scene
Online GPU GiB
Offline GPU GiB
Map B/voxel
Online IoU
Online mIoU
Offline IoU
Offline mIoU
FP32
0.152
4.57
0.319
1.753
3.169
424
67.12
58.40
67.79
59.37
Mixed FP16
0.120
3.60
0.224
1.558
3.252
232
67.08
58.36
67.77
59.35
Table 3: Efficiency and precision comparison on EmbodiedOcc-ScanNet. Runtime and GPU memory are measured on a single RTX 3090.
Figure 6: Real-world robotic platform.
Scene
Travel Dist. (m)
Map Extent (m)
Key Frames
Online Avg. (s/frame)
Online Time (s/scene)
Offline Refine (s)
Online RSS (GiB)
Offline RSS (GiB)
Feature Map (MiB)
Local Cache (MiB)
Reading Room
31.31
10.0×9.0
91
0.630
57.32
4.29
4.172
4.340
35.07
151.55
Laboratory Floor
235.15
57.3×15.6
351
0.764
268.08
25.07
5.136
7.806
215.69
576.50
Table 4: Real-world deployment results in two indoor environments. Map extent denotes the horizontal bounding box of the final stored voxel map. RSS reports the peak process memory usage in each stage, while Local Cache denotes the cached local features stored for offline refinement.
Table 11
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Stage
Per-GPU Batch
Total Batch
Epochs
Training Time (h)
Local Predictor
4 images
16 images
10
24.8
Online Incremental Mapping
1 trajectory (30 frames)
4 trajectories
3
1.3
Offline Global Map Refinement
1 scene map
4 scene maps
30
1.9
Appendix
Table 7: Training configuration and time of the three stages.
Learning rate
Multiplier
IoU
mIoU
4×10−4 (VEOcc; Wang et al., 2026 )
1×
64.55
55.49
8×10−4
2×
66.11
57.35
1.6×10−3
4×
66.90
58.08
2×10−3 (VEOcc++, Ours)
5×
66.98
58.17
Appendix
Table 8: Local learning-rate study on Occ-ScanNet. All results are measured at epoch 10 using the same architecture and remaining training configuration. The best results are marked in bold .
Figure 7: Online deployment scaling on the Laboratory Floor sequence. We track runtime and memory as the global map grows over 351 motion-gated keyframes. The panels show per-keyframe online service latency, current process RSS, and Persistent Voxel Feature Map size, respectively. The centered rolling median over 21 keyframes highlights the latency trend, while the dashed line indicates the 1 s reference.
Figure 8: Qualitative comparison of online incremental mapping on real-world robotic sequences. The first two rows show self-collected sequences previously used in VEOcc, while the last two rows correspond to the two zero-shot deployment sequences reported in the main paper of VEOcc++. For each sequence, we compare the online maps produced by VEOcc and VEOcc++, and show representative local regions in the right column. VEOcc++ yields more complete structures, better continuity of extended objects and surfaces, and improved semantic consistency in the highlighted regions.
Semantic occupancy mapping predicts and maintains 3D maps of occupied, free, and unknown space, labeling objects and surfaces for perception, planning, and navigation.
Crucial for autonomous exploration, online 3D occupancy prediction and mapping incrementally constructs dense spatial representations on the fly. However, recent Gaussian-centric methods struggle with structural boundary fidelity and rely heavily on predefined scene-size priors, fundamentally limiting their operational efficiency. In this work, we present VEOcc, a voxel-centric framework formulated as a recursive perception-and-assimilation paradigm. By eliminating the need for initial scale estimation, VEOcc enables highly streamlined, open-ended map expansion. Furthermore, to robustly aggregate noisy temporal observations within the discrete voxel space, we propose a Spatio-Temporal-Aware Online Update Strategy. It integrates Cross-Temporal Logit Aggregation (TLA) for temporal consistency, Reliability-Aware Confidence Modulation (RCM) for spatial uncertainty calibration, and Confidence-Driven Incremental State Update (CSU) for robust global state assimilation. % Extensive experiments on Occ-ScanNet and EmbodiedOcc-ScanNet demonstrate that VEOcc establishes new state-of-the-art performance in both local and embodied settings, providing an accurate and efficient solution for real-world exploration. Extensive experiments on Occ-ScanNet and EmbodiedOcc-ScanNet demonstrate that VEOcc establishes new state-of-the-art performance in both local and embodied settings. Notably, zero-shot evaluations on self-collected video sequences further confirm its robust out-of-distribution generalization capability in completely unseen real-world environments. Ultimately, our framework provides an accurate and highly efficient solution for autonomous exploration. Code and supplementary visualizations are available on our project page: https://wryzju.github.io/VEOcc/.
Ruoyu Wang, Yong Liu, Sheng Tao +2
Institute of Cyber-Systems and Control, Zhejiang University, Hangzhou, China.
Persistent semantic occupancy mapping is essential for embodied scene understanding. However, perspective-based systems provide limited spatial coverage, while existing panoramic methods primarily predict local volumes from single observations. We introduce PanOVOcc, a training-free framework for persistent open-vocabulary semantic occupancy mapping from panoramic sequences. PanOVOcc unifies panoramic SLAM, open-vocabulary perception, and long-term spatial voxel memory within an online architecture, continuously integrating geometric and semantic evidence into a global, language-queryable map. To facilitate systematic evaluation of this setting, we establish Pan-Replica and Pan-Holo360D, two benchmarks pairing continuous panoramic RGB-D sequences with scene-level semantic occupancy ground truth across synthetic and real-world scenes. Compared with the strongest evaluated baseline for each metric, PanOVOcc improves occupancy IoU and semantic mIoU by absolute +20.03 and +7.06 on Pan-Replica, and by +43.26 and +20.16 on Pan-Holo360D, respectively. The source code and the established benchmarks will be available at https://github.com/bakereet/PanOVOcc.
Di Kuang, Mengfei Duan, Yuhang Wang +2
School of Artificial Intelligence and Robotics, Hunan University, China · National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, China
Embodied agents exploring indoor environments require reliable semantic occupancy memory that persists across observations and revisits. Building such memory is challenging because each observation provides incomplete and uncertain geometric and semantic evidence. We introduce GEM-Occ, a Gaussian Evidence Memory framework that consolidates evidence accumulated over time into persistent semantic occupancy memory. Local predictions are converted into occupied semantic Gaussians and free-space ray evidence. Confidence- and visibility-aware causal updates integrate supporting observations, suppress occupancy contradicted by observed free space, and preserve previously observed structures through occlusion. A hierarchical memory organization supports continued mapping and efficient queries across connected indoor spaces. To evaluate this capability, we introduce HIOcc, a unified benchmark for embodied semantic occupancy memory. HIOcc establishes a shared semantic label space and evaluation framework spanning local prediction, room-level online mapping, and building-level mapping, while accommodating perspective and panoramic observations. Experiments on HIOcc demonstrate that GEM-Occ outperforms existing methods, enabling accurate semantic occupancy prediction and consistent online mapping across spatial scales with efficient memory usage and fast occupancy queries.
Hu Zhu, Bohan Li, Xianda Guo +8
The Hong Kong Polytechnic University · Eastern Institute of Technology · Shanghai Jiao Tong University +3