Abstract
In this work, we consider a multi-wheeled payload transport system where each wheel can be selectively actuated. When not actuated, wheels move freely and do not consume battery power. The system is modeled as an actuated multi-agent system (MAS), with each wheel-motor pair as an agent. Kinematic and dynamic models ensure desired movement, while optimization formulations decide the optimal number and selection of active wheels to conserve battery and reduce motor wear. Our multi-level control framework ensures a near-optimal number of agents are active. Simulation studies demonstrate energy-efficient operation and increased travel distance for the same battery power. Preliminary experimental validation is provided via a custom-built prototype.
Introduction: The Challenge of Over-Actuation
In mobile robotics, "over-actuated" systems—those with more motors than strictly necessary for basic movement—offer significant advantages for heavy lifting and navigating challenging terrains. However, these systems are typically power-intensive. In conventional designs, all motors run simultaneously, often at low-efficiency operating points, leading to rapid battery depletion and uneven mechanical wear.
This research introduces the Multi-Agent Payload Transport Robot (MAPTR), a system designed to manage redundant actuators intelligently. The goal is a "conscious" control system that determines exactly how many motors are required for a specific payload and trajectory, allowing redundant motors to "rest" to maximize overall system lifespan and range.
Methods: Hierarchical Multi-Agent Control
The researchers modeled the robot as a Multi-Agent System (MAS), treating every wheel-motor pair as an individual "agent." To manage these agents, they developed a hierarchical three-level control architecture:
- Level-1 (Agent Velocity Control): A standard PID controller ensuring each active wheel maintains the correct speed.
- Level-2 (Energy Conscious Operation): The system's "brain," utilizing Nonlinear Optimization (Sequential Quadratic Programming). It calculates the optimal number of active motors to maintain peak efficiency based on torque requirements and battery voltage.
- Level-3 (Online Task Allocation): A Linear Integer Program that decides which specific motors to activate. This ensures the workload is distributed evenly across all agents over time, preventing premature wear on any single motor.
The system also incorporates a realistic battery discharge model and accounts for "parasitic" viscous friction caused by dragging inactive, gearless motors.
Key Results and Insights
The core innovation of MAPTR lies in its ability to force active motors into their "sweet spot"—typically 10-20% of stall torque—by deactivating unnecessary units. As shown in simulation results, the system dynamically adjusts the number of active agents based on the payload and operating conditions to maintain high efficiency.
Key findings include:
- Significant Power Savings: Straight-line tests showed up to 31.58% power savings for lighter loads compared to all-motor actuation.
- Extended Range and Operation: For a 13.2kg load, mileage improved from 0.4098 m/J to 0.5347 m/J. With a 30kg payload, the optimized system lasted approximately 800 seconds longer than the non-optimized version.
- Battery-Aware Adaptation: As battery voltage drops, the controller automatically activates more motors to compensate for the loss of individual motor strength, maintaining the desired trajectory.
- Scaling Advantage: Energy savings actually increase as the system scales; a 128-agent system shows a steeper saving curve than a 9-agent system.
Conclusion and Future Work
While the MAPTR demonstrates significant improvements in energy efficiency and wear leveling, the current model assumes flat terrain and no longitudinal wheel slip. Future work will address more complex friction dynamics inherent in "skid-steering" on varied surfaces and introduce holonomic steering for enhanced maneuverability. Additionally, the researchers aim to explore decentralized control laws to move beyond the current locally optimal SQP optimization.
Summary updated