cs.ROMar 29, 2026
SaveContraMap: Contrastive Uncertainty Mapping for Robot Environment Representation
Organizations: School of Computer Science and the Australian Centre for Robotics, University of Sydney, Australia
Abstract
Reliable robot perception requires not only predicting scene structure, but also identifying where predictions should be treated as unreliable due to sparse or missing observations. We present ContraMap, a contrastive continuous mapping method that augments kernel-based discriminative maps with an explicit uncertainty class trained using synthetic noise samples. This formulation treats unobserved regions as a contrastive class, enabling joint environment prediction and spatial uncertainty estimation in real time without Bayesian inference. Under a simple mixture-model view, we show that the probability assigned to the uncertainty class is a monotonic function of a distance-aware uncertainty surrogate. Experiments in 2D occupancy mapping, 3D semantic mapping, and tabletop scene reconstruction show that ContraMap preserves mapping quality, produces spatially coherent uncertainty estimates, and is substantially more efficient than Bayesian kernelmap baselines.
Figures & tables
Fig. 1 : For robust robot operation, scene representations should provide not only spatially consistent mapping but also a measure of uncertainty across the environment. ContraMap augments continuous classification-based mapping with an additional uncertainty output, enabling joint environment representation and direct uncertainty prediction at any queried location. The model reconstructs scene structure while assigning high uncertainty to occluded or weakly observed regions, such as the space behind the table.
Fig. 2 : Predictive uncertainty for a Gaussian Process (GP) and neural-network baselines on three toy datasets. Orange / Magenta points are in-distribution training samples, and Red points are out-of-distribution samples. Background shading indicates relative uncertainty (brighter = higher, darker = lower). Our method best matches the GP “gold standard”: it stays confident near observed data and becomes uncertain as inputs move away from the training distribution.
Fig. 3 : Overview of ContraMap . Our method employs a softmax classifier network with an additional output node to jointly represent the environment and estimate uncertainty. Observed data (e.g., LiDAR scan or segmented point cloud) are augmented with negative samples and noise, where the noise is labeled as an additional -th or ”uncertain” class. The augmented data are then projected into a feature vector space using a set of reference points, forming the training dataset . The model is trained on this dataset using a first-order optimization method. Once trained, we can query the model at arbitrary locations in the environment to obtain either environment mapping results or uncertainty estimates via the additional ”uncertain” node.
| Datasets | Hilbert Maps [ 3 ] | Bayesian Hilbert Maps [ 8 ] | Our | |
| Belgioioso (#features = 8400) | AUC | 0.9904 ( 0.0003) | 0.9901 ( 0.0003) | 0.9901 ( 0.0003) |
| Training time (s) | 2.2609 ( 0.0279) | 722.6140 ( 2.5651) | 3.2367 ( 0.0231) | |
| Inference time (s) | 0.4258 ( 0.0014) | 231.3024 ( 0.0746) | 0.4681 ( 0.0022) | |
| Edmonton (#features = 8800) | AUC | 0.9754 ( 0.0005) | 0.9734 ( 0.0005) | 0.9767 ( 0.0004) |
| Training time (s) | 1.8345 ( 0.0371) | 789.5725 ( 1.7593) | 2.2733 ( 0.0245) | |
| Inference time (s) | 0.6997 ( 0.0016) | 405.7368 ( 0.0842) | 0.7630 ( 0.0011) | |
TABLE I : The performance of our method in comparison to Hilbert Map [ 3 ] and Bayesian Hilbert Map [ 8 ] on the Radish datasets [ 24 ] .
| Metric | HMs | Our |
| mIoU | 0.8096 0.0515 | 0.8412 0.0394 |
| Training Time (s) | 32.6278 1.5994 | 35.9025 1.7746 |
| Inference Time (s) | 203.6160 8.2965 | 203.7221 8.3205 |
TABLE II : Quantitative comparison against 3D Hilbert Maps (HMs) on the SemanticKITTI dataset. Experiments were conducted on the initial scan of 10 distinct sequences. Bold indicates the statistically better result ( ).
| Metric | HMs | Our |
| mIoU | 0.6622 0.1234 | 0.6694 0.1231 |
| Training Time (s) | 32.0929 22.6771 | 35.9777 25.5202 |
| Inference Time (s) | 32.2525 2.6430 | 32.2989 2.6393 |
TABLE III : Quantitative comparison against 3D Hilbert Maps (HMs) on the SceneNet dataset. The evaluation was performed across 60 distinct indoor scenes (rooms). Bold indicates the statistically better result ( ).
Fig. 4 : Occupancy mapping results of each method on Intel dataset.
Fig. 5 : Results extracted from each output of the softmax model.
Fig. 6 : Reconstruction results on SceneNet dataset.
Fig. 7 : Reconstruction results on SemanticKITTI dataset.
| V-PRISM [ 10 ] | Our | |
| IoU | 0.4994 ( 0.0878) | 0.4854 ( 0.0805) |
| Chamfer (m) | 0.0126 ( 0.026) | 0.0124 ( 0.031) |
| Training time (s) | 11.2826 ( 5.6768) | 9.3533 ( 2.1653) |
| Reconstructing time (s) | 229.5563 ( 100.0576) | 3.7560 ( 1.1183) |
TABLE IV : The performance of our proposed method compared to V-PRISM [ 10 ] on YCB dataset.
Fig. 8 : Representative scene reconstructions. Col. 1: RGB. Col. 2: segmented point cloud. Col. 3: our reconstructed mesh. Col. 4: our uncertainty (softmax auxiliary head; lighter = higher). Col. 5: V-PRISM mesh. Col. 6: V-PRISM uncertainty heatmap (lighter = higher). Following [ 10 ] , Cols. 4 and 6 show uncertainty on a 2D scene slice (bottom = nearer the camera). Object boundaries from the top-view projection are overlaid for spatial context.
Fig. 9 : Effect of hinge point count on performance (Intel dataset). Results are shown for our methods with varying numbers of hidden layers and for BHMs .