Lane-level map updating requires temporally consistent, physically valid, and auditable road-scene semantics despite occlusions, ambiguous lane markings, temporary obstacles, and viewpoint changes. We present RoadState, a production-deployed vision-language system that transforms sequential road-scene observations into structured lane-level semantics for Baidu Maps' map-updating pipeline. Its core model, RoadState-VLM, formulates road-scene understanding as autoregressive state estimation, predicting the current structured road state from the previous observation and symbolic state together with the current observation. We introduce Geometric Chain-of-Thought (Geo-CoT), which uses a directed acyclic graph of semantic dependencies as a domain prior to improve semantic consistency and reasoning fidelity across interdependent road-state dimensions. Extensive experiments on a real-world dataset show that RoadState-VLM achieves 94.33% mean accuracy across six road-state dimensions, outperforming the strongest fine-tuned VLM baseline. Ablation studies further demonstrate the complementary effectiveness of Geo-CoT and temporal transition context. A cascaded routing pipeline directs routine cases to lightweight models while reserving complex cases for RoadState-VLM. Since its deployment in March 2025, RoadState has supported lane-level map updating across more than 360 cities and reduced overall production costs by 30%. Together with human-in-the-loop auditing, the system maintains robust performance under stress-test conditions. These results demonstrate how domain-prior-enhanced reasoning, structured state interfaces, selective routing, and human quality control can bridge the gap between VLM-based road-scene understanding and large-scale lane-level map updating.
Figures & tables
Figure 1: An example of structured road information generated from a crowdsourced front-view image sequence.
Figure 2: The RoadState production system. Crowd-sourced and survey imagery is assembled into clips, triaged by complexity, and routed to either lightweight models ( ∼ 85% of frames) or RoadState-VLM ( ∼ 15%). Both branches emit the same structured road state, which must pass a deterministic constraint check before entering map change fusion; violations and low-confidence outputs are escalated to human auditors, whose corrections are recycled as Geo-CoT training data. Section 3 details the RoadState-VLM component.
Figure 3: Overview of the RoadState-VLM framework. The pipeline consists of three main stages covering input and preprocessing for Geo-CoT data construction, the RoadState-VLM training architecture with symbolic feedback injection and a latent reasoning buffer, and autoregressive inference with state evolution for temporally consistent road state estimation.
Model Variant
VRAM (GB)
Latency (s)
FPS
Mean Acc. (%)
Qwen3-VL-4B-Thinking
13.0
22.6
0.088
92.10
RoadState-VLM (8B)
21.8
23.5
0.085
94.33
Table 1: Efficiency–effectiveness comparison between 4B and 8B Thinking variants under the same state-aware paradigm.
Model
Lane Count
Ego-lane Index
Struct Feasibility
Vehicle Feasibility
Obstacle Feasibility
Global Feasibility
Overall
Acc
P
R
Acc
P
R
Acc
P
R
Acc
P
R
Acc
P
R
Acc
P
R
Acc
P
R
RoadSceneBench
-
63.37
65.89
-
75.44
84.67
-
-
-
-
-
-
-
-
-
-
83.83
84.69
-
74.21
78.42
LLaVA-Next-8B
76.96
65.66
49.65
92.66
68.23
55.99
91.14
91.01
90.96
90.51
64.63
65.84
99.79
83.25
66.65
87.22
86.12
86.88
89.71
76.48
69.33
LLaVA-1.5-7B
86.33
73.10
67.09
94.18
54.38
49.06
91.56
91.84
91.02
93.08
73.20
61.69
98.61
55.67
82.68
88.82
87.84
88.48
92.10
72.67
73.34
Qwen2.5-VL-7B
52.49
52.14
44.33
67.26
74.07
64.75
78.19
81.44
75.79
94.22
90.04
60.64
99.83
99.92
66.67
77.34
77.40
72.75
78.22
79.17
64.15
Qwen3-VL-4B-Thinking
90.20
72.85
66.94
92.50
65.71
61.83
91.80
91.42
91.35
93.50
81.76
70.92
99.40
77.31
99.18
85.20
84.06
84.73
92.10
78.85
79.16
Table 2: Quantitative evaluation of RoadState-VLM against state-of-the-art VLMs on the RoadSceneState Benchmark . Performance is measured across six dimensions using Accuracy (Acc), Macro-Precision (P), and Macro-Recall (R); all values are reported as percentages.
Variant
Components
State Dimension Accuracy (%)
Mean Acc. (%)
Geo-CoT
Historical State
Lane Count
Ego Index
Str. Feas.
Veh. Feas.
Obs. Feas.
Global Feas.
M1 (Single-Frame Description)
×
Single-frame
68.69
71.56
-
-
-
61.94
–
M2 (Single-Frame + CoT)
✓
Single-frame
68.10
75.36
79.33
91.31
99.24
77.64
81.83
M3 (State Prediction (Direct))
×
Multi-frame
79.32
82.03
88.48
96.12
99.75
86.54
88.71
M4 (RoadState-VLM)
✓
Multi-frame
91.90
94.60
94.31
95.15
99.70
90.34
94.33
Table 3: Ablation Study on Geo-CoT Reasoning and Historical State Context. Metrics report the Accuracy (Acc) for each state dimension, with all accuracy values reported as percentages. M1 does not output the three feasibility dimensions; therefore, its aggregate mean is omitted and should not be compared directly with M2–M4. M4 represents our full proposed RoadState-VLM framework.
Model Variant
Lane Count
Ego Index
Str. Feas.
Veh. Feas.
Obs. Feas.
Global Feas.
Mean Acc. (%)
Qwen3-VL-Instruct-8B (No CoT)
77.50
81.10
83.20
94.00
99.00
84.20
86.50
Qwen3-VL-Instruct-8B (Geo-CoT)
92.32
94.35
84.89
95.57
99.54
91.01
92.95
Qwen3-VL-Thinking-8B (No CoT)
79.32
82.03
88.48
96.12
99.75
86.54
88.71
RoadState-VLM
91.90
94.60
94.31
95.15
99.70
90.34
94.33
Table 4: Model-variant comparison of Thinking capability and Geo-CoT supervision. All accuracy values are reported as percentages.
Variant
Lane Count
Ego Index
Str. Feas.
Veh. Feas.
Obs. Feas.
Global Feas.
Mean Acc.
w/o Logical Override
88.12
90.30
91.05
92.50
99.12
89.80
91.82
w/o Historical State Feedback (M2)
68.10
75.36
79.33
91.31
99.24
77.64
81.83
RoadState-VLM Full
91.90
94.60
94.31
95.15
99.70
90.34
94.33
Table 5: Accuracy (%) for the Logical Override ablation and the single-frame M2 reference.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Qualitative Case Studies. Representative visualization cases highlighting the complexity of real-world driving scenarios. Red text denotes prediction errors, underscoring the superior reasoning capabilities of our method.
Figure 5: Qualitative Case Studies of Geo-CoT Reasoning. Representative samples illustrating the hierarchical reasoning process of RoadState-VLM across diverse driving environments. Each case demonstrates the model’s logic through Global Observation of static road geometry, Temporal Analysis of ego-motion, and Logical Deduction enforcing axiomatic constraints.
Tsinghua University Baidu Beijing, China · University of Macau Macao, China · Institute of Information Engineering, Chinese Academy of Sciences Beijing, China +2
School of Geospatial Engineering and Science, Sun Yat-Sen University, Zhuhai, 519082, China · Transport Bureau of Nanhai District, Foshan, 528200, China · Laboratory of Intelligent Collaborative Computing, University of Electronic Science and Technology of China, Chengdu, 611731, China +2