New Strategy Boosts Robot Autonomy in Complex Environments
In the rapidly evolving field of robotics, the ability of machines to navigate and map unknown environments autonomously is a critical benchmark for real-world applicability. From search-and-rescue missions in disaster zones to routine inspections in industrial facilities, mobile robots must efficiently explore spaces that are often cluttered, narrow, or expansive. However, traditional exploration algorithms have long struggled with two persistent challenges: the inefficiency of exploration in large, open areas, and the difficulty of navigating through tight, constricted passages. A recent breakthrough from researchers at Beijing University of Technology addresses both issues with a novel compound cooperative strategy that significantly enhances the performance of autonomous mobile robots.
Published in the January 2021 issue of Robot, a peer-reviewed scientific journal known for its contributions to robotics research, the study introduces an innovative approach that combines the strengths of two established methods—Rapidly-exploring Random Tree (RRT) and frontier-based exploration—into a unified framework. The work, led by Li Xiuzhi, He Yaleigh, Sun Yanjun, Zhang Xiangyin, and Zhang Xiaofan from the Faculty of Information Technology and the Engineering Research Center of Digital Community at Beijing University of Technology, demonstrates a marked improvement in exploration efficiency, reducing both time and travel distance while increasing the likelihood of complete environmental coverage.
The limitations of existing methods have long been recognized in the robotics community. Frontier-based exploration, which identifies boundaries between known and unknown regions as potential target points, excels in guiding robots toward unexplored areas. However, in vast, open environments such as large halls or atriums, this method can generate an overwhelming number of candidate points, leading to inefficient navigation and excessive backtracking. On the other hand, RRT, a sampling-based motion planning algorithm, is effective in exploring complex spaces by incrementally building a tree of possible paths. Yet, its random nature makes it particularly slow in environments with narrow entrances, where the probability of generating a node that successfully passes through a tight passage is low.
The research team’s solution lies in a synergistic integration of these two approaches. Their proposed method, termed the “compound cooperative strategy,” leverages RRT for global candidate target point detection to ensure comprehensive mapping of the environment, while simultaneously employing frontier-based detection for local target selection to accelerate exploration. This dual-path system allows the robot to maintain a broad awareness of the entire space while reacting quickly to nearby opportunities for expansion.
At the core of the strategy is a refined process for selecting the optimal next destination. Once candidate points are identified through the combined RRT and frontier mechanisms, the system performs a clustering operation on the global RRT-derived points to reduce redundancy. This step is crucial for computational efficiency, as raw RRT outputs often produce densely packed points that would otherwise require excessive processing. The clustered centers, along with the locally detected frontier points, are then evaluated using a newly designed cost function that balances multiple factors: navigation cost, information gain, and directional alignment.
Navigation cost considers the Euclidean distance from the robot’s current position to the candidate point, as well as the angular adjustment required to face that direction. Information gain quantifies the expected amount of new area that would be revealed upon reaching the point, calculated by measuring the number of unknown grid cells within a fixed radius around the target. Directional alignment ensures that the robot does not waste energy turning sharply unless necessary, promoting smoother and more energy-efficient movement.
To prevent any single factor from dominating the decision-making process, the team implemented a normalization technique that scales each component to a common range before combining them into a final cost value. The point with the highest composite score is selected as the next exploration target. This evaluation method ensures a balanced trade-off between exploration speed and coverage completeness, avoiding the pitfalls of strategies that prioritize short-term gains at the expense of long-term efficiency.
Once the target is chosen, the robot must navigate to it safely. For this, the researchers enhanced the Timed Elastic Band (TEB) algorithm, a dynamic path-planning method known for its ability to generate time-optimal trajectories while avoiding obstacles. Standard TEB can struggle when the target lies behind the robot, potentially generating paths that cut through obstacles due to misaligned initial conditions. The team’s improvement introduces a steering condition constraint that first checks whether the target is within the robot’s forward field of view by analyzing the dot product of directional vectors. If the target is behind, the system recalculates the robot’s yaw angle based on nearby waypoints from the global path, effectively reorienting the robot before local planning begins. This adjustment ensures that the generated path remains collision-free and dynamically feasible.
The modified TEB algorithm also incorporates constraints on linear and angular velocity, acceleration, and minimum clearance from obstacles. These are formulated as penalty functions within a graph optimization framework, allowing the system to balance competing objectives such as speed, safety, and smoothness. By using the g2o library for sparse graph optimization, the implementation achieves real-time performance even on hardware with limited computational resources.
To validate their approach, the team conducted extensive experiments in both simulated and real-world environments. The simulation platform, Gazebo, provided a physically accurate 3D environment where variables could be tightly controlled. In these tests, the robot was tasked with exploring a complex layout featuring multiple rooms connected by narrow corridors—specifically designed to challenge traditional algorithms. The results were striking: the proposed method completed exploration in an average of 275.119 seconds, covering 130.051 meters with only 32 exploration cycles. In contrast, the RRT-based method required 346.967 seconds and 156.734 meters, while the previously developed GTM (Grid-Topological Map) approach failed to complete the task within the allotted time, having consumed 398.872 seconds and traveled 187.586 meters without full coverage.
The real-world experiments were conducted using a custom-built mobile robot equipped with a URG-10LX laser rangefinder, capable of 270-degree scanning at a 10-meter range. The robot operated within a laboratory setting that included open spaces, narrow doorways, and dynamic obstacles. Communication and computation were distributed across a network: the computationally intensive SLAM (Simultaneous Localization and Mapping) module ran on a laptop server, while the robot handled sensor data acquisition, local planning, and motor control. This distributed architecture, built on the Robot Operating System (ROS), ensured robust performance and real-time responsiveness.
In the physical trials, the advantages of the compound strategy became even more apparent. The robot using the new method successfully mapped the entire environment in 1,187.465 seconds, traveling just 97.551 meters over 41 exploration cycles. The GTM method, while eventually covering most of the area, took 1,773.817 seconds and traveled 206.536 meters, remaining incomplete due to excessive backtracking and inefficient pathing. The data clearly show that the proposed method not only reduces exploration time by nearly one-third but also minimizes energy consumption through shorter travel distances.
One of the most compelling aspects of the research is its practical relevance. Unlike deep learning-based approaches that require extensive training data and high-end GPUs, this method relies on deterministic algorithms that can be deployed on standard robotic platforms without specialized hardware. This makes it particularly suitable for applications where computational resources are limited, such as in field robots or low-cost service machines.
Moreover, the system’s modular design allows for easy integration with existing robotic frameworks. The use of ROS ensures compatibility with a wide range of sensors and actuators, while the separation of global and local planning modules enables scalability. Future enhancements could include incorporating dynamic obstacle prediction, multi-robot coordination, or adaptive parameter tuning based on environmental characteristics.
The implications of this work extend beyond academic interest. In industrial automation, for instance, autonomous mobile robots are increasingly used for inventory management in large warehouses. These environments often feature long aisles (open spaces) and narrow access points between storage racks—precisely the conditions where the compound strategy excels. Similarly, in search-and-rescue operations, robots must quickly map collapsed buildings where hallways may be blocked or partially obstructed. The ability to rapidly identify and navigate through narrow passages could mean the difference between life and death.
Another promising application is in smart building maintenance, where robots patrol facilities to monitor structural integrity, HVAC systems, or security. In such scenarios, exploration efficiency directly impacts operational cost and system reliability. A robot that can complete its rounds faster and with fewer energy expenditures translates into longer battery life and reduced downtime.
The research also contributes to the broader discourse on robot autonomy. As robots move from controlled factory floors to unpredictable real-world settings, their ability to make intelligent, context-aware decisions becomes paramount. This study exemplifies how combining classical algorithms in novel ways can yield performance improvements that rival or surpass those of more complex, data-driven models. It underscores the value of algorithmic innovation alongside advances in machine learning.
Looking ahead, the team has laid the groundwork for several potential extensions. One direction involves adapting the cost function to incorporate semantic information—for example, prioritizing exploration of rooms labeled as “kitchen” or “server room” in semi-structured environments. Another possibility is integrating temporal consistency, where the robot learns to predict how obstacles might move over time, further enhancing its navigational intelligence.
The success of this project also highlights the importance of interdisciplinary collaboration. The team brought together expertise in algorithm design, sensor fusion, control theory, and software engineering to create a cohesive system. Their use of open-source tools like ROS and Gazebo not only accelerated development but also ensures that their findings can be replicated and built upon by others in the global robotics community.
In conclusion, the work presented by Li Xiuzhi, He Yaleigh, Sun Yanjun, Zhang Xiangyin, and Zhang Xiaofan represents a significant step forward in autonomous robot exploration. By intelligently combining RRT and frontier-based methods, refining the cost evaluation process, and improving local path planning, they have developed a system that outperforms existing approaches in both simulated and real-world environments. The results demonstrate that thoughtful algorithmic design, rather than sheer computational power, can drive meaningful progress in robotics. As autonomous systems become more prevalent in everyday life, such innovations will play a crucial role in making them more capable, efficient, and reliable.
Li Xiuzhi, He Yaleigh, Sun Yanjun, Zhang Xiangyin, Zhang Xiaofan, Beijing University of Technology, Robot, 10.13973/j.cnki.robot.200009