Simple Machine Learning Model Teaches Robots to Dodge Obstacles in Crowded Spaces

Autonomous mobile robots are increasingly being asked to operate in places where space is at a premium: crowded warehouses, cluttered hospital corridors, busy factory floors and domestic interiors filled with furniture, people and unpredictable clutter. In these dense environments, the difference between a useful robot and a useless one often comes down to a single capability, namely the ability to detect obstacles quickly and decide, in real time, whether to keep going straight, swerve left or swerve right. A new study published in the International Journal of Intelligent Robotics and Applications tackles exactly this problem, and its central claim is surprising: a carefully engineered logistic regression model, of all things, can outperform far more fashionable machine learning approaches when the task is framed correctly.

The research, led by Abhishek Thakur of Birla Institute of Technology Mesra, Jaipur, together with Subhranil Das, Monica Bhutani, Sudhansu Kumar Mishra, Vikash Kumar Gupta and Sitanshu Sekhar Sahu, introduces a model the authors call Adaptive Block Coordinate Descent Logistic Regression, or ABCDLR for short. Rather than treating robot navigation as a continuous control problem requiring heavy computation, the team reframed obstacle avoidance as a three-class classification problem. At every decision point, the robot must choose one of three actions: move forward with no turn, turn left, or turn right. This simplification is the conceptual heart of the work, because it converts a messy, continuous steering problem into a discrete decision that a lightweight statistical model can make in milliseconds.

The sensory setup behind the system is deliberately minimal. Two infrared sensors, mounted to monitor the left and right wheel velocities and their immediate surroundings, work alongside a single ultrasonic sensor that measures the distance to the nearest obstacle ahead. These three data streams, collected in real time, form the input features fed into the classifier. The left and right wheel speeds capture how the robot is currently manoeuvring, while the ultrasonic reading provides the crucial range information that tells the model how urgently a decision is needed. The elegance of this arrangement lies in its cost: infrared and ultrasonic sensors are among the cheapest and most robust ranging devices available, meaning the approach could be deployed on low-budget educational robots and commercial platforms alike without expensive lidar or camera arrays.

Under the hood, ABCDLR is logistic regression trained with an adaptive block coordinate descent optimisation strategy. Block coordinate descent is an iterative optimisation technique in which the parameter vector is partitioned into blocks, and each block is optimised in turn while the others are held fixed. This divide-and-conquer approach can converge more reliably than full gradient methods on ill-conditioned problems, and the adaptive element allows the algorithm to tune its step behaviour as training progresses. The result is a logistic regression model whose coefficients separate the three motion classes cleanly from the sensor data, while remaining interpretable, since each coefficient describes how strongly a given sensor reading pushes the decision toward turning left, turning right or proceeding straight.

To judge whether this simplicity actually pays off, the researchers benchmarked ABCDLR against three widely used machine learning classifiers: K-Nearest Neighbour, Naive Bayes and Gradient Boosting. The evaluation was carried out across three different robot speed conditions, low, medium and high, reflecting the reality that a robot crawling through a crowded corridor faces very different dynamics from one racing across an open floor. Performance was measured using the standard quartet of classification metrics: accuracy, sensitivity, specificity and precision. The authors also interrogated the statistical quality of the fitted logistic regression itself, examining the pseudo R-squared, the Akaike Information Criterion, the Bayesian Information Criterion, the null log-likelihood and the Log-Likelihood Ratio. Together these diagnostics confirm that the model’s fit is statistically meaningful rather than an artefact of overfitting, and they provide a principled basis for comparing model specifications.

The classification results showed that ABCDLR held its own or better against the competing algorithms across the speed regimes, a notable outcome given that Gradient Boosting in particular is typically a formidable benchmark on tabular data. The likely explanation is structural. With only three input features and three output classes, the problem does not reward the capacity of ensemble methods; instead, it rewards a model whose decision boundary is smooth and whose inference is instantaneous. K-Nearest Neighbour, by contrast, must store and search its training data at prediction time, and Naive Bayes rests on independence assumptions that the correlated wheel-speed and distance readings violate. In this narrow, well-posed regime, the disciplined optimisation of a linear classifier wins.

Where the study becomes genuinely compelling is in its second phase: path planning. A classifier that avoids a single obstacle is useful, but a robot must string together hundreds of such decisions to traverse a cluttered space. The team therefore deployed ABCDLR as the core decision engine for navigation through three different types of dense environments, and compared the resulting trajectories against four established path planning approaches: the A* graph-search algorithm, a Fuzzy Logic Controller, the Vector Field Histogram method, and a related logistic regression variant the authors refer to as ASGDLR. These baselines represent the classical canon of robot navigation, from optimal grid search to reactive histogram-based obstacle negotiation, so the comparison is a demanding one.

The reported outcome is that the ABCDLR-driven planner produced competitive or superior navigation performance in the dense test environments, outperforming the four rival approaches. In practical terms, this suggests that a reactive, learning-based decision layer built on a statistically sound classifier can navigate clutter at least as well as methods that require explicit maps, membership functions or histogram maintenance. For warehouse operators weighing the cost of autonomy, the implication is significant: the computational footprint of ABCDLR is small enough to run on modest embedded hardware, while its sensor requirements are trivial compared with vision-based deep learning pipelines that demand GPUs and large labelled datasets.

The work also fits into a broader and rapidly accelerating research conversation. Recent literature on mobile robot navigation spans deep reinforcement learning for unknown environments, convolutional neural networks for vision-based obstacle avoidance, and bio-inspired optimisers such as whale and grey wolf algorithms for path planning. Much of that literature chases ever-greater model complexity. The present study pushes in the opposite direction, arguing that when the perception problem is reduced to a few well-chosen sensor channels and the action space is discretised, classical statistical learning is not merely adequate but advantageous. The authors’ earlier work on collision avoidance in cluttered environments, published in Computers and Electrical Engineering in 2022, laid the groundwork for this conclusion, and the new adaptive optimisation scheme appears to consolidate it.

Caveats remain, and they are worth stating plainly. The study reports that no new datasets were generated or analysed beyond the study itself, and the sensor suite, while cheap, offers a narrow field of perception compared with lidar or stereo vision; obstacles outside the ultrasonic cone remain invisible until the robot turns. The three-class action space also produces reactive rather than deliberative behaviour, which may struggle with dynamic obstacles such as walking people. Nevertheless, the demonstration that a rigorously optimised logistic regression, validated through likelihood-ratio testing and information criteria, can beat A*, fuzzy control and vector field methods in dense environments is a refreshing corrective to complexity inflation in robotics. Sometimes the smartest machine is the one that keeps its model small, its sensors cheap and its statistics honest, and this study makes that case with unusual clarity.

Subject of Research: Machine learning-based obstacle avoidance and path planning for autonomous mobile robots in dense environments

Article Title: Machine learning based intelligent model for obstacle avoidance in dense environments for autonomous mobile robot

Article References: Thakur, A., Das, S., Bhutani, M., Mishra, S. K., Gupta, V. K., & Sahu, S. S. (2026). Machine learning based intelligent model for obstacle avoidance in dense environments for autonomous mobile robot. International Journal of Intelligent Robotics and Applications. https://doi.org/10.1007/s41315-026-00599-8

Image Credits: AI Generated

DOI: 10.1007/s41315-026-00599-8

Keywords: autonomous mobile robot, machine learning, logistic regression, obstacle avoidance, path planning, block coordinate descent, infrared sensors, ultrasonic sensor, collision avoidance, A* algorithm, fuzzy logic controller, vector field histogram