Deploying a language-guided navigation policy in a new building usually means watching it fail. Public-benchmark VLAs degrade scene by scene as layout, appearance, and language drift, and the standard fix of teleoperating fresh trajectories on site does not scale. We replace human demonstration with a single brief iPhone scan. Scan2Instruction compiles that scan into a complete training environment, automatically synthesizing executable navigation episodes with grounded instructions. PRPT-KS then adapts the policy through a closed-form Pareto utility read directly off the reconstructed scene, with no learned critic and no human labels. Across five real environments, the two designs together drive our RealNav to an 85% real-robot success rate, compared with 49% for a RealNav baseline post-trained on human demonstrations, and still achieve 82% in unscanned regions of the same buildings, compared with 41% for the same baseline. Our work turns demonstration-free scene adaptation for language-guided navigation from an aspiration into a working reality.
Figures & tables
Figure 1: One iPhone scan, a deployable language-guided navigation policy. Left: Scan2Instruction compiles a brief handheld scan into scene-specific training episodes, and PRPT-KS adapts RealNav on them with no demonstrations and no human labels. Right: real-robot rollouts following natural-language instructions in the scanned scenes.
Figure 2: Scan2Instruction pipeline. A pretrained MLLM takes rendered images, goal views, and a type-specific prompt to generate Route following, Object Goal, or Human Demand instructions.
Figure 3: PRPT-KS pipeline. A Pareto utility vector read from known-scene geometry drives rollout updates, gates rollouts, and weights preference pairs.
Scene
Area (m 2 )
Scan (min)
Train eps.
Sim test
Real trials
Office
318
20
300
100
60
Cottage
155
13
300
100
40
Teaching
130
11
300
100
40
Apartment 1
90
9
300
100
30
Apartment 2
75
9
300
100
30
Total
768
62
1500
500
200
Table 1: Target real-world scenes and synthesized data. Evaluation instructions are manually checked.
Method
Office
Cottage
Teaching
Apartment 1
Apartment 2
Average
3D scene-graph mapping:
DualMap Jiang et al. (2025)
0.32 / 0.17
0.22 / 0.13
0.17 / 0.10
0.35 / 0.17
0.33 / 0.17
0.28 / 0.15
Navigation VLA:
NaVILA Cheng et al. (2024)
0.25 / 0.20
0.20 / 0.23
0.12 / 0.15
0.19 / 0.20
0.22 / 0.23
0.20 / 0.20
Uni-NaVid Zhang et al. (2024)
0.17 / 0.22
0.22 / 0.23
0.29 / 0.15
0.19 / 0.20
0.21 / 0.20
0.22 / 0.20
StreamVLN Wei et al. (2025)
0.26 / 0.17
0.26 / 0.13
0.10 / 0.13
0.23 / 0.30
0.30 / 0.23
0.23 / 0.18
Table 2: Main evaluation on five scanned scenes. Each cell reports success rate (SR) as sim / real on reconstructed Gaussian scenes and the physical robot under the same task split. Average SR is computed from pooled successes and trials across scenes.
Method
Office
Cottage
Teaching
Apartment 1
Apartment 2
Average
3D scene-graph mapping:
DualMap Jiang et al. (2025)
0.10
0.15
0.20
0.10
0.15
0.14
Navigation VLA:
NaVILA Cheng et al. (2024)
0.15
0.20
0.20
0.10
0.25
0.18
Uni-NaVid Zhang et al. (2024)
0.20
0.30
0.15
0.20
0.25
0.22
StreamVLN Wei et al. (2025)
0.25
0.20
0.30
0.20
0.15
0.22
Table 3: Real-robot evaluation in unscanned regions of the same five scenes as Table 2 : starts, goals, and intermediate waypoints all lie outside the scanned region used for reconstruction, task synthesis, and RL rollout. Each cell reports real-world success rate (SR).
Variant
SR ↑
SPL ↑
NE (m) ↓
OS ↑
RealNav + PRPT-KS (Full)
0.85
0.834
1.42
0.88
RealNav (w/o PRPT-KS post-training)
0.23
0.210
3.50
0.27
w/o behavior-cloning anchor
0.30
0.293
3.49
0.41
w/o Pareto-utility advantage
0.81
0.785
1.31
0.85
w/o Pareto-weighted preference
0.80
0.781
1.32
0.83
Table 4: Ablation analysis in simulation. We compare the full RealNav + PRPT-KS pipeline with RealNav before scene-specific post-training and with variants that individually disable one loss term in the unified post-training objective. Metrics are measured on reconstructed Gaussian scenes.