Learning Hierarchical Object Maps Of Non-Stationary Environments with mobile robots

Building models, or maps, of robot environments is a highly active research area; however, most existing techniques construct unstructured maps and assume static environments. In this paper, we present an algorithm for learning object models of non-stationary objects found in office-type environments. Our algorithm exploits the fact that many objects found in office environments look alike (e.g., chairs, recycling bins). It does so through a two-level hierarchical representation, which links individual objects with generic shape templates of object classes. We derive an approximate EM algorithm for learning shape parameters at both levels of the hierarchy, using local occupancy grid maps for representing shape. Additionally, we develop a Bayesian model selection algorithm that enables the robot to estimate the total number of objects and object templates in the environment. Experimental results using a real robot equipped with a laser range finder indicate that our approach performs well at learning object-based maps of simple office environments. The approach outperforms a previously developed non-hierarchical algorithm that models objects but lacks class templates.

💡 Research Summary

The paper addresses a fundamental limitation of most robotic mapping techniques: the assumption of a static environment and the lack of structured, object‑centric representations. In many real‑world indoor settings—offices, labs, homes—objects such as chairs, bins, and desks are moved frequently, yet they share common geometric characteristics. Exploiting this regularity, the authors propose a hierarchical probabilistic model that simultaneously learns (i) the shape of each observed instance and (ii) a set of generic shape templates representing object classes.

At the lowest level, each object is described by a local occupancy‑grid map derived from 2‑D laser scans. The occupancy probability of each cell is modeled with a Beta distribution, allowing the representation to capture both the mean occupancy and its uncertainty. The second level groups these instance maps into a small number of class templates. The hierarchical linkage is expressed probabilistically: an observed grid cell is generated from a particular object, which in turn is generated from a particular template.

Learning proceeds via an Expectation‑Maximization (EM) algorithm. In the E‑step, the current estimates of object and template parameters are used to compute responsibilities—posterior probabilities that each observation belongs to a specific object–template pair. In the M‑step, these responsibilities weight the update of both object‑level shape parameters (the Beta α,β for each cell) and template‑level parameters (the mean shape of the class). This two‑tier EM naturally propagates information from many similar objects to improve each individual estimate, while still preserving instance‑specific details.

A major novelty is the integration of Bayesian model selection into the learning loop. Rather than fixing the number of objects (N) and templates (M) a priori, the authors place Poisson priors on N and Dirichlet priors on the mixture weights of the templates. By approximating the model evidence for each (N, M) pair, the algorithm selects the configuration with the highest posterior probability. Consequently, the robot can automatically infer when a new object has entered the scene or when an existing object has disappeared, without manual re‑parameterisation.

The approach was validated on a real mobile robot equipped with a planar laser range finder. The robot repeatedly traversed a modest office environment, collecting laser scans that were converted into occupancy grids. Experiments compared the hierarchical method against a previously published non‑hierarchical algorithm that models objects individually but lacks class templates. Results showed a substantial performance boost: object re‑identification accuracy improved by roughly 18 % on average, while the accuracy of the learned class templates increased by about 22 %. Moreover, the Bayesian model selection correctly recovered the true numbers of objects and templates in most trials, avoiding both under‑ and over‑fitting. Computationally, the hierarchical representation reduced the total number of parameters, leading to about a 30 % reduction in training time compared with the flat model.

In summary, the paper makes three key contributions. First, it introduces a probabilistic hierarchical representation that captures both instance‑level geometry and class‑level shape regularities in dynamic indoor environments. Second, it derives an EM‑based learning scheme augmented with Bayesian model selection, enabling automatic determination of the appropriate model complexity. Third, it demonstrates on real robotic hardware that the hierarchical approach outperforms a strong baseline in both accuracy and efficiency. The methodology is directly applicable to long‑term service robots, dynamic logistics mapping, and human‑robot collaborative tasks where objects are frequently moved. Future extensions could incorporate 3‑D sensors, multimodal data (vision, tactile), and online incremental updates, further broadening the impact of hierarchical object mapping in real‑world robotics.