TL;DR
This paper introduces a novel three-layered control architecture for achieving safe coordination among legged robots in dynamic environments. By integrating Control Barrier Functions (CBFs) with forward reachable sets, the system guarantees collision avoidance while enabling robots to follow desired trajectories, even in the presence of moving obstacles and other robots.
What Problem Does This Paper Address?
Coordinating multiple legged robots in dynamic environments, such as those with moving obstacles or other robots, presents a significant challenge. Legged robots are inherently complex due to their high-dimensional, dynamic, and hybrid nature. Ensuring that these robots can safely navigate and collaborate while avoiding collisions, especially when their desired paths intersect or are near dynamic elements, is crucial for applications like search and rescue, logistics, and collaborative manipulation. Existing approaches often struggle to provide robust safety guarantees in such complex, real-world scenarios, particularly when dealing with the interplay of multiple moving agents and unpredictable environments.
Methods
The paper proposes a three-layered control architecture:
- Top Layer: Safety-Critical Coordination Planner: This layer uses Control Barrier Functions (CBFs) based on the forward reachable set to generate safety-ensured trajectories for each robot. It leverages second-order models to predict potential collisions and ensures that the planned paths remain within safe regions.
- Middle Layer: Optimal GRF Trajectory Planner: This layer employs Model Predictive Control (MPC) with Single Rigid Body (SRB) dynamics to compute optimal Ground Reaction Forces (GRFs). The goal is to track the safety-ensured trajectories generated by the top layer while adhering to the robot's dynamics.
- Bottom Layer: Virtual Constraint based Whole-Body Controller: This layer utilizes the full-order robot model and an input-output linearization technique to generate the actual joint torques. It ensures the robot follows the prescribed trajectories and optimal GRFs, maintaining stable locomotion and satisfying friction cone constraints.
The core novelty lies in the integration of forward reachable sets with CBFs for predicting and avoiding collisions in dynamic environments, and the hierarchical decomposition of the control problem across these three layers.
Datasets
Not discussed in the paper. The experiments utilize simulations (RaiSim) and hardware platforms (Unitree quadruped robots) with custom-defined dynamic environments.
Key Insights and Ideas
- Forward Reachable Sets for Predictive Safety: Using forward reachable sets allows the system to predict the future possible states of other agents and obstacles, enabling proactive safety measures rather than reactive avoidance.
- Layered Control for Complexity Management: Decomposing the problem into trajectory planning, GRF generation, and whole-body control simplifies the design and allows each layer to focus on specific aspects of safety and motion.
- CBFs as a Unified Safety Framework: CBFs provide a principled way to encode safety constraints into the optimization problems at different control layers, ensuring that safety is maintained throughout the system.
- Bridging Second-Order and Full-Order Models: The architecture effectively bridges the gap between the simplified second-order models used for high-level planning and the complex full-order dynamics of the physical robots.
Key Results
The paper demonstrates the effectiveness of the proposed approach through both simulation and hardware experiments:
- Simulation: The system successfully coordinated a single agent with three obstacles and four agents, maintaining safe distances and avoiding collisions. Plots show that the distance between agents and obstacles (or other agents) was kept within safe margins.
- Hardware: Experiments with one agent and one obstacle, and two agents, showed that the CBFs remained within the safe set, indicating that coordinated locomotion in dynamic environments tracked safety-ensured trajectories.
Limitations and Future Work
- Limitations:
- The current framework assumes that the states and control laws of other agents are known. This is a simplification for dynamic environments with unknown or unpredictable agents.
- The experiments were conducted in relatively controlled dynamic environments.
- Future Work:
- Extend the framework to handle more complex and uncertain dynamic environments.
- Explore the extension of the safety-critical coordination planner to three dimensions.
- Integrate perception capabilities to allow for varying terrain types and dynamic obstacle detection.
Summary updated