Predictive Control with Indirect Adaptive Laws for Payload Transportation by Quadrupedal Robots

A research team from Virginia Tech, led by Leila Amanzadeh, has published a breakthrough paper in IEEE Robotics and Automation Letters detailing a new hierarchical control framework that significantly enhances the payload-carrying capabilities of quadrupedal robots. The core innovation is an Adaptive Model Predictive Control (AMPC) system that integrates a gradient-descent-based adaptive updating law with a traditional MPC planner. This allows the robot's high-level controller to continuously estimate the unknown mass and dynamics of an attached payload, then use those real-time estimates to plan stable, efficient trajectories. The planned motions are executed by a low-level whole-body controller (WBC).

The system was validated through extensive simulation and hardware experiments, demonstrating remarkable performance gains. On flat terrain, a quadruped equipped with this AMPC framework successfully transported static payloads equivalent to 109% of its own body mass. On rough, uneven experimental terrain, it managed payloads up to 91% of its mass, and could even handle dynamic payloads (like swinging objects) weighing 73% of its mass. The framework proved robust against external push disturbances and obstacles in both indoor and outdoor tests. Performance comparisons showed it significantly outperformed both a standard MPC controller and an L1-adaptive MPC controller, marking a substantial step toward deploying legged robots for real-world logistics and disaster response in unpredictable environments.