CPNav: Calibrated Vision-Language Navigation for Object-Goal Search in Unseen Environments
Authors: Qizhao Chen, Yuanhong Zeng, Shoh Nishino, Anushri Dixit
Organizations: Department of Mechanical and Aerospace Engineering, University of California, Los Angeles, Los Angeles, CA, USA · Department of Electrical and Computer Engineering, University of California, Los Angeles, Los Angeles, CA, USA
Directing a robot to locate a target object in unseen environments remains a central challenge in embodied AI. Vision-Language Model (VLM)-based methods provide rich semantic reasoning but often produce uncalibrated and overconfident action decisions. We introduce CPNav, a navigation framework that combines collision-aware planning with statistically calibrated VLM action scoring for map-free object-goal navigation. From RGB-D input, CPNav constructs a navigability map, generates collision-free motion primitives, and queries two VLMs for goal detection and action scoring. The resulting scores are calibrated offline via conformal prediction (CP) using oracle actions obtained through a tree-structured exploration procedure with backtracking, providing finite-sample guarantees at a user-specified risk level. The CP layer is model-agnostic and can wrap any VLM policy. Online, calibrated VLM scores guide action selection, while a tree-structured exploration history records alternative actions and allows backtracking. We evaluate CPNav in both simulation and real-world hardware settings. In AI Habitat simulation, compared with the uncalibrated baseline method, CPNav achieves relative SR and SPL improvements of 6.2% and 11.4% on HM3D v0.2, and 12.4% and 8.8% on MP3D, respectively. In real-world deployment on a modified Hiwonder MentorPi robot, the corresponding improvements in SR and SPL are 24.9% and 15.2%, respectively. Code is available at https://github.com/practice-lab-ucla/CPnav.
Figures & tables
Fig. 1 : Overview of the CPNav pipeline. (a) The action proposer constructs candidate actions from RGB-D input, prioritizing unexplored regions while enforcing collision safety. (b) The threshold obtained via offline conformal calibration is used to filter out actions whose scores fall below the calibrated bound. (c) The highest-scoring untried action that satisfies the threshold is selected and executed, while the resulting transition is recorded in the exploration tree. (d) Hardware experiments are conducted on a modified Hiwonder MentorPi robot platform.
Fig. 2 : The action proposer uses the RGB-D observation to identify newly discovered regions and generate collision-free candidate actions toward them.
Fig. 3 : Concise stopping and action prompts with their JSON responses. The stopping prompt, fstop(⋅) queries the stopping condition, donet , and GSV, stGSV ; the action prompt, faction(⋅) , queries the per-action confidence score, si,taction . The manually normalized confidence score is multiplied by GSV to form the final per-action score, si,t .
Parameter
Simulation
Hardware
rrobot,rclearance
0.1m,0.05m
Resolution
640×480 RGB-D
Camera Height
0.9m
HFoV / Down Pitch
131∘/0.45rad
101∘/0.4rad
VLM Model
GPT-5 Nano [ 46 ]
Step Limit
200
100
TABLE I : Experimental Parameters
Fig. 4 : Empirical distribution of nonconformity scores Ej used for simulation conformal calibration. The dashed red line indicates the calibrated nonconformity quantile q^=0.940 , evaluated at the Beta-adjusted quantile level 1−α^ .
Policy
HM3D
MP3D
SR (%) ↑
SPL (%) ↑
Dist. (m)
SR (%) ↑
SPL (%) ↑
Dist. (m)
No Calibration [ 18 ]
81.0
40.3
35.4
68.5
43.0
36.4
Simple Set
76.7
39.5
32.8
71.0
44.3
33.4
Prompt Set
79.7
38.7
37.4
71.5
45.3
31.6
Ensemble Set
74.0
39.1
28.0
66.5
41.8
28.2
STRIVE [ 47 ]
82.5
41.5
17.6
71.5
42.3
21.0
TABLE II : Simulation evaluation on HM3D and MP3D (200 episodes per benchmark). Distance is reported in meters.
Variant
SR (%)
SPL (%)
Mean Distance (m)
CPNav ( q^=0.940 )
86.0
44.9
29.2
No Calibration
81.0
40.3
35.4
Height = 0.45 m
67.0
31.4
44.6
Height = 1.70 m
79.3
40.5
30.4
HFoV = 69 ∘
40.0
18.7
44.8
HFoV = 101 ∘
72.7
30.5
40.9
TABLE III : Parameter sensitivity analysis in simulation.
Policy
SR (%)
SPL (%)
Mean Distance (m)
No Calibration
66.7
42.0
9.0
CPNav
83.3
48.4
6.3
TABLE IV : Hardware results.
Fig. 5 : Floor plans of the two hardware evaluation environments.
Policy
50 Steps (25%)
100 Steps (50%)
200 Steps (100%)
SR (%)
SPL (%)
SR (%)
SPL (%)
SR (%)
SPL (%)
No Calibration
60.3
37.7
75.7
39.8
81.0
40.3
CPNav
73.0
43.2
80.0
44.4
86.0
44.9
TABLE V : Performance under restricted step budgets.
Fig. 6 : Top-down exploration maps comparing the No Calibration baseline and CPNav on an HM3D plant-search episode. The baseline repeatedly revisits explored regions and terminates unsuccessfully after 44 steps, whereas CPNav reaches the target plant in 18 steps with fewer revisits and less redundant exploration.