1Shenzhen Key Laboratory of Robotics Perception and Intelligence, Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, China.
2Jiaxing Research Institute, Southern University of Science and Technology, Jiaxing, China.
3Peng Cheng National Laboratory, Shenzhen, China.
Abstract: Precise and flexible cart-pushing is a challenging task for mobile robots. The motion constraints during cart-pushing and the robot's redundancy lead to complex motion planning problems, while variable payloads and disturbances present complicated dynamics. In this work, we propose a novel planning and control framework for flexible whole-body coordination and robust adaptive control. Our motion planning method employs a local coordinate representation and a novel kinematic model to solve a nonlinear optimization problem, thereby enhancing motion maneuverability by generating feasible and flexible push poses. Furthermore, we present a disturbance rejection control method to resist disturbances and reduce control errors for the complex control problem without requiring an accurate dynamic model. We validate our method through extensive experiments in simulation and real-world settings, demonstrating its superiority over existing approaches. To the best of our knowledge, this is the first work to systematically evaluate the flexibility and robustness of cart-pushing methods in experiments.
The motion planning experiments are conducted in PyBullet, where a dual-arm robot pushes a cart to follow a reference trajectory defined by waypoints. To isolate the performance of the motion planning algorithms from contact dynamics, the cart's URDF is omitted. We formulate the local planning problem as an optimization problem—which is the most common and reasonable approach in the field of mobile manipulation—and compare our custom local planner model (TT model, denoted as *) and improved model (LF model, denoted as **) against typical local planner models from other works. Due to limited maneuverability, other models show larger tracking errors and even divergence, while ours remain flexible and robust.
NMPC
WB-MPC
TT-MPC*
LF-MPC**
NMPC
WB-MPC
TT-MPC*
LF-MPC**
NMPC
WB-MPC
TT-MPC*
LF-MPC**
The control experiments evaluate the controller’s ability to accurately execute the motion planner’s commands under varying payloads and external disturbances. Due to the cart’s large inertia and significant uncertainty in its dynamics (large variance in the dynamic parameters), conventional controllers struggle to eliminate steady-state error or introduce severe delays, leading to overshoot or oscillation. We design a disturbance rejection controller in the local coordinate frame, which achieves superior control performance compared to both PD controllers and typical adaptive controllers.
The commands (desired orientation) are set to fluctuate within ±0.2 rads, while a 22.2 kg payload is applied as a disturbance to evaluate the controllers' adaptability.
To evaluate the proposed local planning and control framework in practical scenarios, we design a set of challenging navigation tasks—pushing the cart through narrow spaces, which is common in supermarkets or dense crowds. We generate global paths, represented as sparse waypoints, using Hybrid A* to guide the cart through these constrained environments, and track the waypoints using 3 baseline frameworks. The sharp reference trajectories resulting from frequent posture adjustments in tight spaces. The accumulation of planning and control errors leads to task failure for baseline methods, whereas our framework completes the tasks successfully, demonstrating superior maneuverability and robust control performance.
Xiao's model+PD controller [4] ✘
Xiao's model (arms fixed in a central pose). The local planner fails to effectively track the sharp global path, resulting in hazardous accelerations due to the large turning radius.
Schulze's model+PD controller [8] ✘
Schulze's model (direct whole-body planning in local coordinates). Due to the coupling of optimization objectives, local planning exhibits unnatural oscillations and failed.
LF model+adaptive controller (Ours)✔
Our model complete the process of recognizing a cart, grasping the cart, and maneuvering the cart through a narrow corridor.
Xiao's model+PD controller [4] ✘
Failed due to large turning radius.
Schulze's model+PD controller [8] ✘
Failed due to abnormal swinging.
LF model+adaptive controller (Ours)✔
Succeed.
Transportation task with one arm failure
Impact resistance thanks to nonlinear control law
Joint PD controller
The joint PD controller cannot provide sufficient stiffness during impacts, causing the cart to accelerate to speeds that the manipulator cannot control, resulting in self-collision
Ours
Our proposed controller rapidly increases stiffness at the moment of impact to stabilize the system, then maintains compliance without significant overshoot once the impact disappears