TL;DR
This paper introduces a novel layered control framework for enabling two quadrupedal robots to cooperatively navigate complex environments. By developing a reduced-order model and proposing both centralized and distributed Model Predictive Control (MPC) strategies, the authors demonstrate robust and efficient cooperative locomotion, significantly outperforming baseline approaches.
What Problem Does This Paper Address?
Cooperative locomotion for legged robots, especially when constrained by physical connections (like carrying a shared payload), presents a significant challenge. Unlike simpler multi-robot systems (like wheeled robots or aerial vehicles), legged robots are inherently unstable, high-dimensional, and subject to complex dynamics and contact constraints. Existing methods often struggle with real-time planning for these complex systems, particularly when dealing with the "curse of dimensionality" in centralized approaches or the communication overhead in distributed ones. This paper tackles the problem of enabling two holonomically constrained quadrupedal robots to move together robustly and efficiently, even on challenging terrains and under disturbances.
Methods
The authors propose a layered control architecture:
- High-Level Trajectory Planning (MPC):
- Reduced-Order Model: They introduce an interconnected Single Rigid Body (SRB) dynamics model. This model captures the essential dynamics of two robots linked together, including the interaction wrenches (forces and torques) between them, which is crucial for cooperative tasks.
- Centralized MPC: A single MPC controller optimizes the trajectories for both robots simultaneously, considering the interconnected SRB dynamics and holonomic constraints.
- Distributed MPC: To reduce computational load, they develop a distributed MPC where each robot's controller optimizes its own local state and control inputs. This approach uses a one-step communication delay and an agreement protocol to share information and reach a consensus on optimal solutions.
- Low-Level Whole-Body Motion Control (Nonlinear Controllers):
- Mapping to Full-Order Dynamics: Distributed nonlinear controllers are used to translate the optimal reduced-order trajectories and ground reaction forces (GRFs) generated by the MPCs into actual joint torques for each robot.
- QP and Virtual Constraints: These controllers leverage quadratic programming (QP) and the concept of virtual constraints (desired behaviors that simplify complex dynamics) to ensure the robots' full-order dynamics accurately track the planned trajectories.
Datasets
Not discussed in the paper. The paper focuses on developing and validating a control framework through simulations and experiments with physical robots.
Key Insights and Ideas
- Interconnected SRB Model is Key: The authors demonstrate that simply treating robots independently or using simpler models (like LIP) is insufficient for robust cooperative locomotion when holonomic constraints are present. Capturing the interaction wrenches between the robots in a reduced-order model is critical.
- Distributed MPC Achieves Similar Performance with Less Computation: A major finding is that the distributed MPC approach achieves performance comparable to the centralized MPC in terms of stability and robustness, but with a significant reduction in computation time (up to 70%). This makes real-time cooperative control more feasible.
- Layered Approach Bridges the Gap: The layered architecture effectively bridges the gap between the simplified reduced-order models used for planning and the complex full-order dynamics of the robots. The low-level controllers successfully impose the planned motions onto the physical robots.
- Robustness to Uncertainty is Crucial and Achieved: The proposed framework shows remarkable robustness to various uncertainties, including unknown payloads, rough terrains, and external disturbances, which is essential for real-world applications.
Key Results
- Simulation:
- Nominal MPC (without considering holonomic constraints) failed to stabilize cooperative locomotion.
- Both centralized and distributed MPCs successfully stabilized cooperative locomotion on simulated rough terrains with unknown payloads, achieving success rates of 54.2% and 53.8%, respectively, compared to 0% for the nominal MPC.
- Similar robustness was observed under simulated external forces and payloads, with success rates of 55.6% and 53.9% for centralized and distributed MPCs, respectively.
- Experiments:
- The layered control approach was successfully validated on two physical quadrupedal robots (Al robots) in various indoor and outdoor scenarios.
- Experiments included locomotion on rough terrain, slippery surfaces, and under external disturbances (pushes and pulls), with unknown payloads.
- The system demonstrated robust tracking of desired trajectories and feasible GRFs across these challenging conditions.
- The distributed MPC consistently showed significantly reduced computation time compared to the centralized MPC.
Limitations and Future Work
- Limitations:
- Small Planning Horizon: The current MPC planning horizon is relatively small (25 ms), limiting the ability to perform highly dynamic or agile maneuvers.
- Simplified Holonomic Constraints: The paper primarily assumes a simple distance constraint between interaction points. More complex constraints (e.g., restricting rotations) were explored but could be further integrated.
- Two-Agent Focus: The current framework is demonstrated for two agents. Extending it to more agents presents computational challenges for centralized approaches and requires careful handling of communication and consensus in distributed ones.
- Sensor Noise: Experimental estimations of global coordinates rely on kinematic estimators, which can introduce noise and affect coordination.
- Future Work:
- Longer Planning Horizons: Investigating longer horizons for MPC, possibly with Piecewise Affine (PWA) optimal control formulations to handle stance leg changes.
- Sophisticated Constraints & Loco-Manipulation: Developing algorithms for more complex cooperative tasks, such as those involving robotic arms for loco-manipulation, and addressing more intricate holonomic constraints.
- Multi-Agent Extension: Extending the framework to handle cooperative locomotion for three or more agents, potentially using reinforcement learning or data-driven techniques.
- Robust State Estimation: Developing algorithms for robust estimation of global agent coordinates in the presence of noisy sensor measurements.
Summary updated