Abstract
Legged robots exhibit significant potential across diverse applications, including but not limited to hazardous environment search and rescue missions and the exploration of unexplored regions both on Earth and in outer space. However, the successful navigation of these robots in dynamic environments heavily hinges on the implementation of efficient collision avoidance techniques. In this research paper, we employ Collision Cone Control Barrier Functions (C3BF) to ensure the secure movement of legged robots within environments featuring a wide array of static and dynamic obstacles. We introduce the Quadratic Program (QP) formulation of C3BF, referred to as C3BF-QP, which serves as a protective filter layer atop a reference controller to ensure the robots' safety during operation. The effectiveness of this approach is illustrated through simulations conducted on PyBullet.
Introduction
Legged robots are prized for their ability to navigate complex, uneven terrain where wheels fail. However, making them move safely in dynamic environments—where people or other robots might cross their path—is a massive challenge. Most high-level controllers for legged robots (like Convex MPC or ZMP-based walking) are designed primarily for stability and gait generation, not necessarily for reactive obstacle avoidance. If an obstacle suddenly appears, these controllers might not react in time, or their reaction might cause the robot to fall. This research addresses the need for a computationally efficient, real-time "safety layer" that can override a robot's desired path only when a collision is imminent, ensuring the robot stays safe while still trying to achieve its original goal.
Methodology
The core of the approach is the Collision Cone Control Barrier Function (C3BF) formulated as a Quadratic Program (QP).
Control Barrier Functions (CBFs)
These are mathematical tools used to define a "safe set" of states. As long as a specific inequality is satisfied, the system is guaranteed to stay within that safe set.
Collision Cones
This geometric concept defines a set of relative velocities between the robot and an obstacle that would lead to a collision. If the robot's velocity vector stays outside this "cone," it won't hit the object.
The C3BF-QP Filter
The authors combine these into a "safety filter." The robot has a "desired" command from its baseline controller (e.g., "walk forward at 1 m/s"). The C3BF-QP looks at this command; if it's safe, it passes it through. If it leads to a collision, the QP finds the minimal possible change to that command to keep the robot safe.
Dimensional Decomposition
To keep the math simple and fast, avoidance is split into two planes:
- Vertical Obstacles (x-y plane): Modeled using unicycle dynamics for steering around obstacles.
- Horizontal Obstacles (x-z plane): Modeled to allow the robot to change its height (ducking).
Experimental Setup
This research validated the approach using the PyBullet physics simulation engine. The algorithms were tested on two specific robot models:
- A Quadruped (four-legged) robot using a Convex MPC baseline controller.
- A Bipedal (two-legged) robot using a Zero Moment Point (ZMP) preview controller.
Key Insights
- The "Minimal Intervention" Principle: The C3BF-QP acts as a "protective filter." It doesn't take over the whole job of walking; it only nudges the acceleration or velocity commands when a collision is predicted. This allows it to work on top of almost any existing walking controller.
- Elliptical Obstacle Approximation: By treating obstacles as ellipses and the robot as a point (padded by its own width), the authors created a robust geometric framework that handles various obstacle shapes.
- Ducking as a Safety Maneuver: While most collision avoidance research focuses on steering left or right, this paper explicitly derives a CBF for the x-z plane. This allows legged robots to utilize their unique morphology to duck under obstacles.
Results
The C3BF-QP was successfully integrated with two very different baseline controllers (MPC for quadrupeds and ZMP for bipeds) without requiring a redesign of the underlying gaits.
In simulations, robots successfully overtook static obstacles and navigated around moving obstacles by adjusting their heading. The quadruped successfully identified horizontal bars and lowered its body height to pass underneath them, regaining its original height once clear. Furthermore, the authors provided formal proofs (Theorems 1 and 2) demonstrating that their proposed functions are indeed valid Control Barrier Functions, ensuring theoretical safety guarantees.
Conclusion and Future Work
The current results are strictly in simulation (PyBullet). Real-world sensors (LiDAR, cameras) introduce noise and latency that weren't fully explored here. Additionally, navigating through a dense crowd or a highly cluttered room remains a challenge for future study. The authors suggest that "Neural CBFs" using machine learning could be a promising way to handle even more complex, non-linear scenarios in the future.
Summary updated