A new probabilistic semantic SLAM approach improves trajectory and map accuracy, strengthening robustness against perceptual aliasing and classifier errors

In environments with similar objects, autonomous robots must determine whether a detected object is a known landmark or a newly encountered one while mapping. To mitigate associated limitations, a research team developed BPDA-GMM, a Bayesian framework, that handles both data association and landmark creation simultaneously using accumulated evidence. Tests demonstrate that the approach enhances trajectory accuracy and object mapping while maintaining real-time performance on embedded hardware.

Reliable navigation, place recognition, and object interaction require autonomous robots to maintain precise environmental maps. Semantic simultaneous localization and mapping (SLAM) add meaning to a robot's map by representing landmarks as recognizable objects rather than only as geometric points. By adding contextual meaning into spatial maps, semantic SLAM enables robots to successfully execute complex downstream tasks like autonomous inspection, asset tracking, and human-robot assistance.

A major challenge is deciding which mapped object produced a new observation. If several objects belong to the same category, such as multiple chairs, a robot must determine whether a new detection belongs to one of those existing objects or represents a previously unseen object. Existing probabilistic data-association approaches can assume a fixed number of landmarks, require repeated computation as a map grows, or depend on manually tuned settings for deciding when a new landmark should be created. A recent publication in IEEE Robotics and Automation Letters introduces Bayesian Probabilistic Data Association via Gaussian Mixture Models (BPDA-GMM), which formulates data association and new-object creation as a joint probabilistic inference problem.

Professor Nak Young Chong at the Japan Advanced Institute of Science and Technology (JAIST) in Japan, along with Thanh Nguyen Canh, a doctoral student at JAIST, and other team members, developed BPDA-GMM for semantic SLAM. Their findings were published online on August 21, 2026.

"We wanted the robot to treat object association as an evolving probabilistic decision rather than a sequence of separate yes-or-no choices. By accumulating evidence over time and assigning probability to both existing and new landmarks, our approach can make the object-level map more stable as the environment becomes more complex," explains Prof. Chong. 

The team's central idea is to use a statistical model known as Dirichlet-process to keep a running record of how strongly observations support existing landmarks. These counts are updated incrementally, allowing the robot to carry forward information from its previous observations without having to recalculate everything from the beginning. At the same time, the likelihood of creating another new object automatically changes as the map grows. This helps prevent duplicate registrations in increasingly crowded maps.

For each semantic detection, BPDA-GMM first narrows the possible matches using both object-class and geometric information. It then combines the likelihood of each candidate with the Bayesian prior to calculate association probabilities. Object landmarks are represented as semantic Gaussian distributions, which together form a Gaussian mixture model. 

When evidence is ambiguous, an additional α-divergence tempering step can sharpen the association decision. The back-end also decouples semantic landmark refinement from direct pose updates, helping prevent noisy detections from corrupting the robot's estimated trajectory.

Experiments in simulation and on a real indoor sequence showed that BPDA-GMM improved trajectory accuracy, semantic mapping quality, and robustness to perceptual aliasing and classifier errors compared to conventional methods. The improvements were particularly pronounced in difficult outdoor simulations where correctly matching objects was challenging. The best conventional method had a median position error of about 29.57 meters, whereas BPDA-GMM reduced this to 8.15 meters. In the indoor experiment, BPDA-GMM correctly mapped 77 of the 84 ground-truth objects and achieved an F1 score of 0.749. By comparison, one competing approach produced 101 mapped objects, creating multiple duplicate entries. 

Any robot that must maintain a reliable object level map over long periods can benefit from this approach. Home-assistance robots that remember furniture and appliance locations, warehouse and factory transport robots, inspection drones, and autonomous platforms that map objects such as parked cars, poles, and trees can all benefit from this framework. 

"Our goal is to make object-level maps reliable enough for robots to use over long periods and across practical settings. Because these maps are understandable to both people and robots, more reliable association could support spoken instructions, shared maps across robot fleets, and autonomous systems. Any robot that must maintain a reliable object level map over long periods can benefit from this approach," explains Prof. Chong.

By reducing duplicate landmarks and improving the handling of ambiguous observations, BPDA-GMM offers a way to make semantic maps more stable while retaining real-time operation on embedded hardware. The researchers identify future directions including richer multi-modal object representations, open-vocabulary semantics, active planning for resolving ambiguous associations, and integration into multi-robot systems.

Figure 1.

Image Title: BPDA-GMM system overview
Image Caption: Schematic illustration showing how each semantic detection is first filtered by a semantic-geometric gate, then assigned CRP-weighted association probabilities. These weights update semantic Gaussian landmarks in the front-end, while the dominant mixture component is converted into a max-mixture semantic factor for the decoupled back-end.
Image Credit: Prof. Nak Young Chong from Japan Advanced Institute of Science and Technology, Japan
Image Source link: https://doi.org/10.1109/LRA.2026.3726389
License Type: CC BY 4.0
Usage Restrictions: Credit must be given to the creator.

Figure 2.

Image Title: Qualitative object maps on the indoor sequence (Left: GPDA, Middle: SlideSLAM, Right: BPDA-GMM)
Image Caption: Schematic illustration showing BPDA-GMM yields one compact landmark per object, while baselines over-instantiate duplicates or miss classes.
Image Credit: Prof. Nak Young Chong from Japan Advanced Institute of Science and Technology, Japan
Image Source link: https://doi.org/10.1109/LRA.2026.3726389
License Type: CC BY 4.0
Usage Restrictions: Credit must be given to the creator.

Reference

Additional information for EurekAlert

Funding information

The research was supported by JST SPRING, Japan (Grant Number JPMJSP2102) and Asian Office of Aerospace Research and Development (Grant Number FA2386-25-1-4034).

September 01, 2026