Consensus-based optimization (CBO): Towards Global Optimality in Robotics

Zero-order optimization has recently received significant attention for designing optimal trajectories and policies for robotic systems. However, most existing methods (e.g., MPPI, CEM, and CMA-ES) are local in nature, as they rely on gradient estimation. In this paper, we introduce consensus-based optimization (CBO) to robotics, which is guaranteed to converge to a global optimum under mild assumptions. We provide theoretical analysis and illustrative examples that give intuition into the fundamental differences between CBO and existing methods. To demonstrate the scalability of CBO for robotics problems, we consider three challenging trajectory optimization scenarios: (1) a long-horizon problem for a simple system, (2) a dynamic balance problem for a highly underactuated system, and (3) a high-dimensional problem with only a terminal cost. Our results show that CBO is able to achieve lower costs with respect to existing methods on all three challenging settings. This opens a new framework to study global trajectory optimization in robotics.

💡 Research Summary

This paper introduces Consensus‑Based Optimization (CBO) as a globally convergent, zero‑order method for robotic trajectory and policy optimization, addressing the limitations of widely used local sampling techniques such as Model Predictive Path Integral control (MPPI), the Cross‑Entropy Method (CEM), and Covariance Matrix Adaptation Evolution Strategy (CMA‑ES). The authors first recast these existing methods within the framework of estimation‑of‑distribution algorithms (EDAs), showing that they all rely on a parametrized Gaussian search distribution whose mean is updated by a soft‑max weighting of sampled trajectories. While this approach provides a form of cost smoothing, it remains fundamentally local: the Gaussian shape restricts exploration, especially in high‑dimensional control spaces where finite sample sizes cannot faithfully approximate the true cost landscape, leading to the well‑known “curse of dimensionality” and premature convergence to sub‑optimal minima.

CBO departs from this paradigm by abandoning an explicit parametric distribution. Instead, a population of particles directly represents the search distribution. Each particle evolves according to the stochastic differential equation

(du_i = -\lambda (u_i - \bar u),dt + \sigma |u_i - \bar u|,dW_i),

where (\bar u) is the consensus point defined as a cost‑weighted average of all particles, (\lambda) controls the drift toward consensus, and (\sigma) scales a multiplicative noise term that preserves diversity. This dynamics yields a non‑parametric, potentially highly asymmetric distribution with long tails in promising directions, allowing the algorithm to explore globally while still concentrating mass around low‑cost regions. Under mild assumptions on (\lambda) and (\sigma), prior work has proven almost‑sure convergence of the empirical particle distribution to the global minimizer as the number of particles tends to infinity.

The authors validate CBO on three challenging robotic benchmarks: (1) a long‑horizon trajectory optimization for a simple 2‑DOF system, (2) a dynamic balance task for a highly under‑actuated double‑pendulum, and (3) a high‑dimensional humanoid posture optimization with only a terminal cost. In the first scenario, MPPI and CEM remain trapped in local minima despite increasing sample counts, whereas CBO’s consensus point steadily approaches the global optimum as particle count grows. In the second scenario, the non‑linear contact dynamics and under‑actuation cause CMA‑ES to diverge or converge extremely slowly; CBO quickly discovers a balanced trajectory by leveraging inter‑particle consensus and noise‑driven exploration. In the third, the high dimensionality renders Gaussian‑based methods ineffective due to insufficient sampling; CBO’s irregular, heavy‑tailed distribution still identifies a low‑cost posture, achieving significantly lower final costs than all baselines with comparable computational budgets.

Overall, the paper demonstrates that CBO offers three key advantages over traditional zero‑order methods: (i) genuine global exploration without reliance on a fixed‑shape Gaussian, (ii) a flexible, non‑parametric representation that can adapt its shape to the problem’s landscape, and (iii) theoretical guarantees of convergence that scale with particle number. The authors also discuss how CBO subsumes other meta‑heuristics (e.g., particle swarm, differential evolution) within a unified diffusion‑based framework, opening avenues for real‑time policy updates, hardware‑in‑the‑loop experiments, and extensions to learning‑based control. The work positions CBO as a promising new tool for achieving global optimality in complex robotic systems where gradient information is unavailable or unreliable.