Where to look and how to move? A robot navigating an unmapped environment must do both at once, and the two goals pull against each other. The regions most worth observing are the ones the map knows least about, and those are exactly where the robot cannot trust its collision margins. We resolve this tension by introducing Splat-CBF, an active perception control barrier function that steers the camera toward the next best view while collision avoidance is enforced as a hard constraint. Safety is enforced by a risk-aware control barrier function that turns the Average Value-at-Risk of the Gaussian field into a single smooth hard constraint. Perception is enforced by a second barrier that rewards camera orientations with high expected Fisher information gain near the robot's planned path. The two meet in a quadratic program where safety is hard and perception is soft, with a slack penalty that adapts to how often perception has already been relaxed and how close the robot is to an uncertain region. We verify the method in indoor simulations, a Isaac Kinova manipulator and in experiments on an Ackermann-drive robot. Our results assert that robot navigates faster, gathers more information, and runs faster online than safety-only and perception-only baselines, giving up informative motion only when safety requires it.
Figures & tables
Fig. 1: System overview. Each RGB-D observation updates the 3DGS map and nominal plan, while a unified CBF-QP combines a hard safety barrier with a soft trajectory-aware perception barrier to generate safe, informative control actions in closed loop.
Scene
Method
Safety [%] ↑
Success [%] ↑
Time [ms] ↓
Min. dist.
Stonehenge 1.16×105 Gs
PolyMerge
95
96
1.17
0.55
SAFER-Splat
100
97
33.33
0.47
Splat-CBF (Ours), ε=.30
100
99
1.01
0.73
Splat-CBF (Ours), ε=.10
100
77
1.00
0.98
Statues 2.02×105 Gs
PolyMerge
99
100
1.16
2.80
SAFER-Splat
99
98
47.16
1.73
TABLE I: Safety-navigation comparison across four 3DGS scenes. Time denotes the mean online control computation per step.
Fig. 2: Qualitative comparison of safety-navigation trajectories in a 3DGS environment: Splat-CBF (ours, ), SAFER-Splat ( ), and PolyMerge ( ).
Fig. 3: Ablation of the perception-slack weighting using adaptive, fixed, and hard formulations. The ◊ denotes the robot pose at t=68 s, when the adaptive method reaches the goal. Insets show cumulative uncertainty reduction over time, with the adaptive result in gray as reference.
Fig. 4: The Splat-CBF trajectory (left) Green ( ) denotes compatible safety and perception objectives ( δp=0 ), while red ( ) indicates conflicts resolved by relaxing perception ( δp>0 ) while preserving hard safety. Compared with CAAP and passive mapping, Splat-CBF reaches the goal fastest while maintaining positive mean safety and perception barriers.
Fig. 5: Qualitative trajectory comparison in the same 3DGS environment. From the common start (green), Splat-CBF (ours, ) and CAAP ( ) reach the goal (red), whereas Safety Only ( ) terminates before reaching it.
Fig. 6: Closed-loop safety–perception arbitration on the Ackermann robot. The safety barrier hs remains positive, while the perception constraint is selectively relaxed through δp during conflicts; the adaptive weight ξ regulates the cost of this relaxation online.
Fig. 7: Online 3DGS reconstruction during the Ackermann experiment (left) and the corresponding map uncertainty (right), highlighting poorly observed regions relevant to safe-active navigation.
Fig. 8: Real-robot Ackermann experiment illustrating safety–perception conflict resolution. The dashed white curve denotes the nominal trajectory; Green segments ( ) indicate compatible safety and perception objectives, while red segments ( ) denote perception relaxation near obstacles, where the hard safety constraint takes priority.
Fig. 9: Kinova manipulator setup for safe active perception. (a) Simulated Kinova Gen3 operating in the laboratory shelf environment with the corresponding online Immediate3DGS rendering and selected camera views. (b) Kinova arm registered in the reconstructed 3D Gaussian map, showing the planned camera motion and next-best-view poses.
Selector
Controller
PSNR ↑
SSIM ↑
Depth RMSE [m]↓
Minimum henv↑
Median QP Solve Time [ms] ↓
Random
Safety Only
15.78
0.649
0.413
0.126
0.94
Splat-CBF
16.85
0.682
0.381
0.136
1.13
COVER
Safety Only
17.08
0.712
0.376
0.089
0.91
Splat-CBF
17.96
0.737
0.355
0.138
1.15
FisherRF
Safety Only
16.69
0.667
0.401
0.077
0.96
Splat-CBF
16.87
0.681
0.393
0.132
1.18
TABLE III: Reconstruction, safety, and computational performance for different view selectors and controller variants.