Hierarchical Motion Control for Autonomous Wheelchairs moving in a Crowded Environment

Hierarchical Motion Control for Autonomous Wheelchairs moving in a Crowded Environment∗ Elisa Gaetan1,2, Mohammad Jeddi1,2, Alessandro Mattioli1, Paolo Falcone1and Laura Giarr´ e1 EXTENDED ABSTRACT Autonomous systems, such as vehicles and robots, are rapidly advancing, and their ability to navigate complex and dynamic environments is crucial for practical deployment. This need is particularly evident in assistive mobility devices, such as wheelchairs, where safety, autonomy, and adaptability are paramount [1]. We introduce a hierarchical planning framework that integrates both a macro planner and a micro planner for navigating robots in crowded environments [2]. The macro planner provides high-level guidance by identifying preferred regions of the environment based on macroscopic features, while the micro planner operates at a finer scale, accounting for dynamic constraints and safety considerations to compute feasible short-term trajectories. This design enables the robot to act proactively by anticipating possible deviations in the predicted trajectories of surrounding obstacles. As a result, the robot avoids collisions through smoother, less abrupt motion adjustments. As a concrete application, we propose the design of an autonomous wheelchair capable of safely and efficiently navigating both structured and dynamic environments. The system is divided into two main layers: low-level motor control and high-level motion planning. This architecture allows for fine-grained control of movement while also enabling intelligent decision-making in the presence of obstacles. By embedding safety-driven algorithms into the control system, the wheelchair is not only responsive to real-time changes in its environment but also adaptive to unexpected scenarios, thereby ensuring that users can maintain independence without compromising safety. High-Level Motion planning We design a hierarchical motion planner, specifically a twolayer controller, that combines a macroscopic and a microscopic planner in cascade (Fig.1). In the macroscopic planner, we propose to leverage global features of pedestrian groups—such as density and kinetic energy—to capture their collective behavior. Unlike several approaches in the literature that rely on fluid dynamics, particlebased models, or potential fields [3], [4], [5], [6], which often *This work was partly supported by the EU Next-Generation (PNRR) within and partly by the Mobility National Center, SPOKE 6 and under the Program of National Interest PRIN 2022 - Control of Assistive Robots in crowded Environments (CARE) - CUP E53D2300046 0006. 1E. Gaetan, M. Jeddi, L. Giarr´ e and P. Falcone are with the Engineering Department ’Enzo Ferrari’, University of Modena and Reggio Emilia, Modena, Italy. {elisa.gaetan, mohammad.jeddi, laura.giarre, paolo.falcone}@unimore.it 2E. Gaetan and M. Jeddi are also with the Italian National PH.D. DAUSY, Politecnico di Bari, Bari, Italy. {elisa.gaetan, m.jeddi}@phd.poliba.it depend on strong assumptions (e.g., access to pedestrians’ final goals or cost functions), our method adopts a more lightweight and assumption-free formulation. Specifically, we compute macroscopic quantities—density, kinetic energy, and entropy—to build distinct spatial layers of the environment. By combining these layers with appropriately chosen weights, we assign a score to each cell in the map, thereby identifying the most suitable regions for the autonomous agent to traverse (Fig.2, dashed line). Once the macro planner identifies the preferred region, the micro planner is responsible for reaching the target while avoiding collisions with obstacles. These obstacles can be either static or dynamic, and in some cases, may form groups. While the goal of reaching the target can be straightforwardly incorporated into the planning framework, guaranteeing collision avoidance becomes significantly more complex, especially when a prediction horizon is considered. In the context of MPC, a prediction horizon is used to compute the optimal control actions over future time steps. However, in crowded or pedestrian-dense environments, accurately predicting human movements becomes extremely difficult, especially when assumptions about the pedestrians’ final destinations are removed. To address this challenge, we propose modeling the direction of motion of pedestrians—represented as unicycle systems—using a probability distribution with finite support. This distribution is incorporated into the optimization problem via slack variables introduced in the collision avoidance constraints. In essence, when the predicted position of a pedestrian corresponds to a high probability, the controller behaves more conservatively. Conversely, less likely positions lead to more relaxed constraints, allowing for less conservative behavior. Within this framework, the robot is not constrained by a fixed set of collision avoidance constraints; instead, it operates under a set of flexible constraints that adapt according to the specific scenario. This flexibility in the safety constraints allows the robot to reach its target more efficiently, as it can partially relax its caution with respect to certain moving obstacles. Such adaptability is especially important in specific robotic applications, such as autonomous wheelchairs. In these scenarios, it would be unreasonable to require the robot to always take longer paths merely to avoid influencing the motion of nearby pedestrians. Since the robot is carrying a person, it would be unfair to impose unnecessary delays on them solely for the convenience of others. However, this mechanism can adapt to different situations. For instance, if two obstacles are observed to be moving closer together—suggesting the potential formation of a cohesive group—the planner is more likely to propose an external 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629644 61 Perception Layer Macroscopic Planner Microscopic Planner Environment Macro features Cell-based motion planning Control action Fig. 1: Motion planner block scheme. path, avoiding navigation between them. Conversely, if two obstacles are observed to be moving apart, this may indicate that there is space to pass between them, as they appear to exhibit flexible behavior. In other words, the degree of flexibility in the collision avoidance constraints can vary depending on the scenario and on predictions of obstacle motion. Autonomous Wheelchair Application The motion planner described in the previous section is implemented in closed loop for an autonomous wheelchair application, as shown in Fig. 3. The wheelchair control system employs a hierarchical structure with two nested loops. The outer position loop minimizes distance and heading errors relative to the planned trajectory. Feedback for this loop is obtained through an integrated sensor suite, including GPS, Pozyx, LiDAR, and an IMU. These sensors provide real-time measurements to an Extended Kalman Filter (EKF), which performs sensor fusion and delivers accurate estimates of the wheelchair’s position and orientation. Using this information, a PID controller generates the desired linear and angular velocities, which are converted into reference wheel speeds. The inner velocity loop tracks these wheel speed references using encoder feedback. A low-level PID controller regulates the velocity error and outputs the required PWM signals to the motors. Operating at a higher frequency, this loop ensures fast disturbance rejection, while the outer loop guarantees accurate trajectory tracking. Early experiments with the low-level motor control framework, using odometry feedback, show promising outcomes. The platform successfully maintains consistent speed profiles and executes basic manoeuvres such as straight-line motion, -3 -2 -1 0 1 2 3 4 -10 -8 -6 -4 -2 0 2 4 Fig. 2: Robot trajectory considering the outcomes of the macroscopic planner (dashed line). Fig. 3: Block diagram of the closed-loop system turning, and simple trajectory following. These preliminary results confirm the effectiveness of the control architecture and establish a reliable foundation for subsequent high-level control and planning tasks. Our current efforts focus on system integration and incremental real-world testing. Initial tests are performed in structured indoor settings, gradually extending to outdoor and semi-structured environments where challenges such as uneven terrain, variable lighting, and dynamic obstacles are present. REFERENCES [1] C. Mavrogiannis, F. Baldini, A. Wang, D. Zhao, P. Trautman, A. Steinfeld, and J. Oh, “Core Challenges of Social Robot Navigation: A Survey,” ACM Transactions on Human-Robot Interaction, vol. 12, no. 3, pp. 1–39, Sep. 2023. [2] E. Gaetan, P. Falcone, A. M. Bayen, and L. Giarr´ e, “Soft Control Barrier Functions for autonomous robots moving in human populated environments,” (in preparation for IEEE T-SMC). [3] Y. Sun and H. Liu, “Crowd evacuation simulation method combining the density field and social force model,” Physica A: Statistical Mechanics and its Applications, vol. 566, p. 125652, Mar. 2021. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S037843712030950X [4] Y. Jiang, B. Chen, X. Li, and Z. Ding, “Dynamic navigation field in the social force model for pedestrian evacuation,” Applied Mathematical Modelling, vol. 80, pp. 815–826, Apr. 2020. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S0307904X19306043 [5] X.-X. Jian, S. Wong, P. Zhang, K. Choi, H. Li, and X. Zhang, “Perceived cost potential field cellular automata model with an aggregated force field for pedestrian dynamics,” Transportation Research Part C: Emerging Technologies, vol. 42, pp. 200–210, May 2014. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S0968090X14000321 [6] H. Liang, J. Du, and S. Wong, “A Continuum model for pedestrian flow with explicit consideration of crowd force and panic effects,” Transportation Research Part B: Methodological, vol. 149, pp. 100–117, Jul. 2021. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S0191261521000849 62