ADMM-Based Distributed MPC with Control Barrier Functions for Safe Multi-Robot Quadrupedal Locomotion

TL;DR

ADMM-based distributed MPC with control barrier functions for quadrupedal robots reduces planning time by 51%.

cs.RO 🔴 Advanced 2026-03-20 47 views
Yicheng Zeng Ruturaj S. Sambhus Basit Muhammad Imran Jeeseop Kim Vittorio Pastore Kaveh Akbari Hamed
distributed control model predictive control control barrier functions multi-robot systems quadrupedal robots

Key Findings

Methodology

This paper proposes a distributed model predictive control (MPC) framework based on the alternating direction method of multipliers (ADMM) with control barrier function (CBF) constraints for safe trajectory planning in multi-robot quadrupedal systems. By introducing a node-edge splitting technique with consensus constraints, the method decomposes the global problem into independent node-local and edge-local quadratic programs that can be solved in parallel using neighbor-to-neighbor communication.

Key Results

  • Result 1: In a four-agent scenario, the average per-cycle planning time is reduced by approximately 51%, achieving performance comparable to centralized MPC.
  • Result 2: Hardware experiments validate the method's effectiveness in uncertain environments, ensuring safe inter-robot distances.
  • Result 3: Numerical simulations demonstrate the method's dynamic feasibility and safety when navigating rough terrain and external disturbances.

Significance

This research provides an efficient distributed solution for safe trajectory planning in multi-robot systems, particularly in the field of quadrupedal robots. By integrating CBFs into the MPC framework, it addresses the explicit coupling issue among robots, enhancing system robustness and real-time capabilities. This framework not only improves computational efficiency but also opens new possibilities for robot collaboration in complex environments.

Technical Contribution

Technical contributions include: 1) a novel node-edge splitting technique allowing fully decentralized trajectory optimization; 2) integration of CBFs into the MPC framework providing safety guarantees; 3) distributed solution of non-convex problems via ADMM, significantly reducing computational burden.

Novelty

This paper is the first to combine CBFs with ADMM for distributed MPC in multi-robot quadrupedal systems, offering a novel decomposition method that significantly enhances computational efficiency and safety assurance.

Limitations

  • Limitation 1: In extremely complex environments, the conservativeness of CBFs may lead to overly cautious trajectory planning.
  • Limitation 2: The algorithm's convergence may be limited under specific non-convex conditions.

Future Work

Future research directions include: 1) optimizing CBF parameters to enhance flexibility; 2) extending to other types of robotic systems; 3) exploring more efficient distributed optimization algorithms to further reduce computation time.

AI Executive Summary

In recent years, quadrupedal robots have made significant strides in navigating complex terrains, enabling applications in inspection, search and rescue, and collaborative transportation. However, extending quadrupedal locomotion to multi-robot teams, while improving efficiency and robustness, introduces significant challenges. These challenges primarily arise from high-dimensional dynamics, hybrid contact behaviors, and inter-agent safety constraints.

Model predictive control (MPC) has emerged as a powerful framework for trajectory generation and control, systematically handling system dynamics and constraints while optimizing performance objectives. Despite its success in single-robot systems, MPC faces explicit coupling issues in multi-robot systems, particularly in quadrupedal robots.

This paper proposes a distributed MPC framework based on the alternating direction method of multipliers (ADMM) with control barrier function (CBF) constraints for safe trajectory planning in multi-robot quadrupedal systems. By introducing a node-edge splitting technique with consensus constraints, the method decomposes the global problem into independent node-local and edge-local quadratic programs that can be solved in parallel using neighbor-to-neighbor communication. This distributed structure not only reduces computational burden but also enhances robustness to communication failures, supporting real-time implementation.

The effectiveness of the proposed approach is demonstrated through hardware experiments on two Unitree Go2 quadrupedal robots and numerical simulations involving up to four robots navigating uncertain environments. The results show that the distributed formulation achieves performance comparable to centralized MPC while reducing the average per-cycle planning time by approximately 51% in the four-agent case, indicating improved computational efficiency while maintaining trajectory quality and safety.

Despite these advancements, the method may face challenges in extremely complex environments where the conservativeness of CBFs could lead to overly cautious trajectory planning. Additionally, the algorithm's convergence may be limited under specific non-convex conditions. Future research directions include optimizing CBF parameters to enhance flexibility, extending to other types of robotic systems, and exploring more efficient distributed optimization algorithms to further reduce computation time.

Deep Analysis

Background

In recent years, quadrupedal robots have made significant strides in navigating complex terrains, enabling applications in inspection, search and rescue, and collaborative transportation. However, extending quadrupedal locomotion to multi-robot teams, while improving efficiency and robustness, introduces significant challenges. These challenges primarily arise from high-dimensional dynamics, hybrid contact behaviors, and inter-agent safety constraints. Model predictive control (MPC) has emerged as a powerful framework for trajectory generation and control, systematically handling system dynamics and constraints while optimizing performance objectives. Despite its success in single-robot systems, MPC faces explicit coupling issues in multi-robot systems, particularly in quadrupedal robots.

Core Problem

The core problem in multi-robot systems is achieving efficient trajectory planning while ensuring safety. Specifically, explicit coupling among robots and complex environmental constraints make centralized MPC difficult to scale. Additionally, computational complexity and communication requirements grow rapidly with the number of robots, limiting the feasibility of real-time applications. Therefore, developing an MPC framework that can operate efficiently in a distributed environment is crucial.

Innovation

The core innovations of this paper include: 1) a novel node-edge splitting technique allowing fully decentralized trajectory optimization; 2) integration of CBFs into the MPC framework providing safety guarantees; 3) distributed solution of non-convex problems via ADMM, significantly reducing computational burden. Unlike previous distributed MPC methods, this approach explicitly reformulates the centralized safety-critical MPC problem using a structured distributed optimization framework, enabling the global problem to be solved in parallel through neighbor-to-neighbor communication.

Methodology

  • �� Propose a distributed MPC framework based on ADMM with control barrier function (CBF) constraints for safe trajectory planning in multi-robot quadrupedal systems.
  • �� Introduce a node-edge splitting technique with consensus constraints to decompose the global problem into independent node-local and edge-local quadratic programs.
  • �� These subproblems can be solved in parallel using neighbor-to-neighbor communication, enabling fully decentralized trajectory optimization.
  • �� Integrate the framework into a hierarchical locomotion control architecture for quadrupedal robots, combining high-level distributed trajectory planning, mid-level nonlinear MPC, and low-level whole-body control.

Experiments

The experimental design includes hardware experiments on two Unitree Go2 quadrupedal robots and numerical simulations involving up to four robots in uncertain environments. Hardware experiments validate the method's effectiveness in complex environments, including obstacle avoidance and robustness to external disturbances. Numerical simulations demonstrate the method's dynamic feasibility and safety when navigating rough terrain and external disturbances. Simulations are conducted using the RaiSim physics engine, while hardware experiments are performed in an indoor laboratory environment.

Results

Experimental results show that the distributed framework reduces the average per-cycle planning time by approximately 51% in the four-agent scenario, achieving performance comparable to centralized MPC. Additionally, hardware experiments validate the method's effectiveness in complex environments, ensuring safe inter-robot distances. Numerical simulations demonstrate the method's dynamic feasibility and safety when navigating rough terrain and external disturbances.

Applications

This method can be directly applied to safe trajectory planning in multi-robot systems, particularly in the field of quadrupedal robots. Its distributed structure reduces computational burden and enhances robustness to communication failures, supporting real-time implementation. Potential industry impacts include improving robot collaboration in complex environments, enhancing applications in search and rescue, inspection, and transportation.

Limitations & Outlook

Despite significant advancements in improving computational efficiency and safety, the method may face challenges in extremely complex environments where the conservativeness of CBFs could lead to overly cautious trajectory planning. Additionally, the algorithm's convergence may be limited under specific non-convex conditions. Future improvements include optimizing CBF parameters to enhance flexibility, extending to other types of robotic systems, and exploring more efficient distributed optimization algorithms to further reduce computation time.

Plain Language Accessible to non-experts

Imagine you're in a kitchen cooking a meal. You need to handle multiple tasks simultaneously, like chopping vegetables, frying, and boiling soup. Each task requires your attention and coordination, and you want to finish everything in the shortest time possible. Now, imagine you have several friends helping you, each responsible for one task, but they need to communicate and coordinate at certain steps to ensure the whole process runs smoothly. This is similar to the multi-robot system in this paper, where each robot is like a friend handling its own task (trajectory planning) but needs to communicate with others to ensure safety (avoiding collisions). Control barrier functions (CBFs) are like safety rules in the kitchen, ensuring everyone works within a safe distance. By using a distributed approach, each robot can work independently but can achieve overall coordination through simple communication, just like you and your friends efficiently working together in the kitchen.

ELI14 Explained like you're 14

Imagine you're playing a game with your friends on the playground. Each of you has a task, like one person passing the ball and another shooting. To avoid bumping into each other, you need to communicate while playing to make sure everyone has enough space. This is like the robot system in this paper, where each robot has its task (like moving to a place) but needs to communicate to ensure safety (not bumping into each other). Control barrier functions (CBFs) are like the rules of the game, ensuring everyone plays within a safe distance. By using a distributed approach, each robot can act independently but can achieve overall coordination through simple communication, just like you and your friends playing a game on the playground. Isn't that cool?

Glossary

Model Predictive Control (MPC)

MPC is a control strategy that optimizes current control inputs by predicting future system behavior. It can handle multiple constraints and dynamic changes.

In this paper, MPC is used to generate motion trajectories for quadrupedal robots.

Control Barrier Function (CBF)

CBF is a mathematical tool used to ensure system safety by restricting system states to avoid unsafe situations.

In this paper, CBF is used to ensure safe distances between robots in a multi-robot system.

Alternating Direction Method of Multipliers (ADMM)

ADMM is an algorithm for distributed optimization that improves computational efficiency by decomposing problems and solving subproblems in parallel.

In this paper, ADMM is used to implement distributed MPC.

Node-Edge Splitting

This is a decomposition technique that breaks down a global problem into independent node-local and edge-local problems for parallel solving.

In this paper, this technique is used to decompose trajectory optimization problems in multi-robot systems.

Quadrupedal Robot

A quadrupedal robot is a robot with four legs, capable of stable movement over complex terrains.

In this paper, experiments were conducted using Unitree Go2 quadrupedal robots.

Distributed Optimization

Distributed optimization is a method of solving optimization problems through collaboration among multiple computing nodes.

In this paper, distributed optimization is used for trajectory planning in multi-robot systems.

Nonlinear Model Predictive Control (NMPC)

NMPC is an MPC method that considers nonlinear system dynamics, providing higher accuracy.

In this paper, NMPC is used for mid-level control to track high-level references.

Whole-Body Control (WBC)

WBC is a control strategy that considers the full-body dynamics of a robot to achieve complex motion tasks.

In this paper, WBC is used for low-level control to track full-body dynamics.

RaiSim Physics Engine

RaiSim is an efficient physics engine for simulating interactions between robots and physical environments.

In this paper, RaiSim is used for numerical simulation experiments.

Unitree Go2 Robot

The Unitree Go2 is a quadrupedal robot with high mobility and stability, suitable for various applications.

In this paper, the Unitree Go2 is used for hardware experiment validation.

Open Questions Unanswered questions from this research

  • 1 Open Question 1: How can CBF parameters be optimized in extremely complex environments to enhance flexibility and adaptability? Current methods may be too conservative, limiting robot movement flexibility.
  • 2 Open Question 2: How can ADMM algorithm convergence be ensured under non-convex conditions? Existing methods may not guarantee convergence under specific conditions.
  • 3 Open Question 3: How can the distributed MPC framework be effectively scaled to larger multi-robot systems? Current computational complexity may limit system scalability.
  • 4 Open Question 4: How can more efficient communication protocols be implemented in distributed environments to reduce latency and data loss? Existing communication mechanisms may affect system real-time performance.
  • 5 Open Question 5: How can uncertainty in dynamic environments be better handled in multi-robot systems? Current methods may not fully address rapidly changing environmental conditions.

Applications

Immediate Applications

Search and Rescue Missions

Multi-robot systems can be used in post-disaster search and rescue, achieving efficient search and collaboration through distributed MPC, ensuring safe distances between robots.

Industrial Inspection

In complex industrial environments, quadrupedal robots can be used for equipment inspection and maintenance, improving efficiency and safety through distributed control.

Collaborative Transportation

Multi-robot systems can be used for collaborative transportation tasks, achieving efficient material handling and path optimization through distributed trajectory planning.

Long-term Vision

Smart City Management

In the future, multi-robot systems can be used for smart city management and maintenance, achieving efficient resource scheduling and environmental monitoring through distributed control.

Automated Agriculture

In agriculture, multi-robot systems can be used for automated planting and harvesting, improving agricultural production efficiency and resource utilization through distributed optimization.

Abstract

This paper proposes a fully decentralized model predictive control (MPC) framework with control barrier function (CBF) constraints for safety-critical trajectory planning in multi-robot legged systems. The incorporation of CBF constraints introduces explicit inter-agent coupling, which prevents direct decomposition of the resulting optimal control problems. To address this challenge, we reformulate the centralized safety-critical MPC problem using a structured distributed optimization framework based on the alternating direction method of multipliers (ADMM). By introducing a novel node-edge splitting formulation with consensus constraints, the proposed approach decomposes the global problem into independent node-local and edge-local quadratic programs that can be solved in parallel using only neighbor-to-neighbor communication. This enables fully decentralized trajectory optimization with symmetric computational load across agents while preserving safety and dynamic feasibility. The proposed framework is integrated into a hierarchical locomotion control architecture for quadrupedal robots, combining high-level distributed trajectory planning, mid-level nonlinear MPC enforcing single rigid body dynamics, and low-level whole-body control enforcing full-order robot dynamics. The effectiveness of the proposed approach is demonstrated through hardware experiments on two Unitree Go2 quadrupedal robots and numerical simulations involving up to four robots navigating uncertain environments with rough terrain and external disturbances. The results show that the proposed distributed formulation achieves performance comparable to centralized MPC while reducing the average per-cycle planning time by up to 51% in the four-agent case, enabling efficient real-time decentralized implementation.

cs.RO math.OC

References (20)

Safety-Critical and Distributed Nonlinear Predictive Controllers for Teams of Quadrupedal Robots

B. Imran, Jeeseop Kim, Taizoon Chunawala et al.

2025 3 citations ⭐ Influential View Analysis →

A Nonlinear MPC Framework for Loco-Manipulation of Quadrupedal Robots With Non-Negligible Manipulator Dynamics

Ruturaj S. Sambhus, Kapi Ketan Mehta, A. Sadeghi et al.

2025 2 citations ⭐ Influential View Analysis →

Architectures for distributed and hierarchical Model Predictive Control - A review

R. Scattolini

2009 1329 citations

Representation-Free Model Predictive Control for Dynamic Motions in Quadrupeds

Yanran Ding, Abhishek Pandala, Chuanzheng Li et al.

2020 174 citations View Analysis →

Per-Contact Iteration Method for Solving Contact Dynamics

Jemin Hwangbo, Joonho Lee, Marco Hutter

2018 248 citations

Discrete-Time Control Barrier Function: High-Order Case and Adaptive Case

Y. Xiong, Dihua Zhai, M. Tavakoli et al.

2022 85 citations

Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers

2011 5132 citations

Distributed predictive control: A non-cooperative algorithm with neighbor-to-neighbor communication for linear systems

M. Farina, R. Scattolini

2012 274 citations

FATROP: A Fast Constrained Optimal Control Problem Solver for Robot Trajectory Optimization and Control

Lander Vanroye, A. Sathya, J. Schutter et al.

2023 44 citations View Analysis →

Decentralized control of complex systems

D. Siljak

2012 956 citations

Autonomous navigation of underactuated bipedal robots in height-constrained environments

Zhongyu Li, Jun Zeng, Shuxiao Chen et al.

2021 34 citations View Analysis →

Dynamic Locomotion in the MIT Cheetah 3 Through Convex Model-Predictive Control

J. Carlo, Patrick M. Wensing, Benjamin Katz et al.

2018 733 citations

Robust Predictive Control for Quadrupedal Locomotion: Learning to Close the Gap Between Reduced- and Full-Order Models

Abhishek Pandala, Randall T. Fawcett, Ugo Rosolia et al.

2022 43 citations

Stability and optimality of distributed model predictive control

Aswin N. Venkat, J. Rawlings, Stephen J. Wright

2005 292 citations

Centroidal dynamics of a humanoid robot

D. Orin, A. Goswami, Sung-Hee Lee

2013 474 citations

Optimization-Based Control for Dynamic Legged Robots

Patrick M. Wensing, Michael Posa, Yue Hu et al.

2022 238 citations View Analysis →

Distributed Nonlinear Model Predictive Control for Connected Vehicles Trajectory Tracking and Collision Avoidance with Ellipse Geometry

Mohamed Abdelaal, S. Schön

2019 2 citations

Distributed Non-Convex Model Predictive Control for Non-Cooperative Collision Avoidance of Networked Differential Drive Mobile Robots

R. Mao, Hu Dai

2021 10 citations

Walking in Narrow Spaces: Safety-Critical Locomotion Control for Quadrupedal Robots with Duality-Based Optimization

Qiayuan Liao, Zhongyu Li, A. Thirugnanam et al.

2022 43 citations View Analysis →

Multirobot Adversarial Resilience Using Control Barrier Functions

Matthew Cavorsi, Lorenzo Sabattini, Stephanie Gil

2024 24 citations