---
title: 'MARS-ADMM: Trajectory & Manifold Optimization'
url: https://www.emergentmind.com/topics/mars-admm
type: topic
---

# MARS-ADMM: Trajectory & Manifold Optimization

MARS-ADMM refers to two distinct but related algorithms based on the alternating direction method of multipliers (ADMM), both designed for challenging settings in optimization. Originally introduced as "Multi-robot ADMM with Robustness and Stiffness Decoupling" for multi-robot trajectory optimization [2111.07016], it has recently appeared as the first single-loop Stochastic Riemannian ADMM for nonsmooth composite optimization on manifolds [2512.22750]. Both variants leverage structured ADMM splitting, but target different domains—trajectory planning and general composite optimization respectively.

## 1. Formulation and Problem Settings

The first MARS-ADMM, in multi-robot trajectory optimization [2111.07016], seeks a solution to a constrained nonconvex optimization problem involving robot trajectories. The primary variables are a sequence of Bézier control points (for UAVs) or configuration variables $\theta\in\R^{\mathrm{DOF}\times N}$ (for robot arms), alongside a uniform time step $\Delta t>0$. The objective functional is a sum of segment-wise costs $\mathcal{O}(X_i,\Delta t)$, potentially including jerk, snap, or time penalties. Collision avoidance is enforced by requiring separating planes between all robot and obstacle convex hulls, with additional kinematic and dynamic constraints:
\[
\|V_i\|\le v_{\max}\Delta t,\quad \|A_i\|\le a_{\max}\Delta t^2,
\]
where $V_i$ and $A_i$ denote velocity and acceleration for trajectory segment $i$. All hard constraints are recast via log-barrier penalties $-\gamma\log(g(\cdot))$ of finite weight, resulting in a fully coupled unconstrained Lagrangian.

The stochastic Riemannian MARS-ADMM [2512.22750] generalizes the setting to composite nonsmooth problems on a Riemannian manifold $\M\subset\mathbb{R}^n$:
\[
\min_{x\in\M}\;F(x) + g(Ax),
\]
where $F(x)=\mathbb{E}_{\xi}[f(x,\xi)]$ is a smooth (possibly nonconvex) stochastic objective, $g$ is convex and possibly nonsmooth, and $A$ is a linear operator. The problem can be equivalently rewritten in block-separable form by introducing an auxiliary variable $y$:
\[
\min_{x\in\M, y\in\mathbb{R}^m} F(x) + g(y)\quad \text{s.t.} \quad Ax-y=0.
\]

## 2. ADMM Splitting and Stiffness Decoupling

In the multi-robot setting [2111.07016], MARS-ADMM exploits the distinct numerical properties of cost terms vs. constraint terms. Log-barrier constraints are characterized as stiff due to unbounded gradients near the feasibility region, while the per-segment trajectory cost remains well-conditioned. The algorithm introduces per-segment slacks $\bar{X}_i$ and $\Delta\bar{t}_i$, enforcing $X_i(\theta)=\bar{X}_i$ and $\Delta t=\Delta\bar{t}_i$ via an augmented Lagrangian with penalty parameter $\varrho$:
\[
\mathcal{L}_{\rm AL} = \sum_i [\mathcal{O}(\bar{X}_i,\Delta\bar{t}_i)] - \gamma\sum_{\text{constraints}}\log(\cdot) + \text{quadratic penalties and dual terms}.
\]
ADMM alternation consists of:
- (A) Minimizing stiff terms with respect to $(\theta,\Delta t)$,
- (B) Fast, unconstrained minimization of smooth cost with respect to slacks $(\bar{X}_i,\Delta\bar{t}_i)$,
- (C) Dual ascent for multipliers,
- (D) Alternating minimization on plane parameters.

The stochastic Riemannian MARS-ADMM [2512.22750] avoids double-loop complexity by integrating all updates into a single loop. It uses momentum-based variance-reduced gradient estimators (SARAH-style recursion) exclusively for the smooth component and leverages constant-size batch sampling.

## 3. Algorithmic Structure and Key Innovations

The original MARS-ADMM algorithm [2111.07016] does not update its penalty parameter $\varrho$, mirroring classical ADMM. Stiffness decoupling—keeping log-barrier penalties strictly in the $(\theta, \Delta t)$ subproblem—ensures that slack updates and trajectory cost minimization remain numerically stable and efficient. Separating-plane normals are kept unit-length via exponential map parameterization.

In the stochastic Riemannian context [2512.22750], each iteration involves:
- Penalty parameter update: $\rho_k = c_\rho k^{1/3}$,
- $y$-update via proximal mapping on $g$,
- Stochastic estimator formation and Riemannian retraction for $x$-update,
- Dual variable and step size updates with adaptive schedules,
- Momentum recursion on gradient estimator, implementing a Riemannian SARAH-style variance reduction.

The momentum-based gradient estimator is:
\[
v_{k+1} = \nabla f_{\mathcal{S}_{k+1}}(x_{k+1}) + (1-\alpha_{k+1}) \mathcal{T}_{x_{k+1}, x_k}(v_k - \nabla f_{\mathcal{S}_{k+1}}(x_k)),
\]
with $\alpha_{k+1} = c_\alpha k^{-2/3}$.

## 4. Theoretical Guarantees

The multi-robot MARS-ADMM [2111.07016] exhibits three guarantees:
- **Feasibility and Collision-Avoidance**: Iterates are maintained strictly feasible by performing line-search shrink-back using continuous-collision-detection, ensuring no constraint violation at convergence.
- **Homotopy Preservation**: With a strictly feasible initialization, the trajectory cannot cross any obstacle boundaries, so the homotopy class of the path is preserved.
- **Convergence**: Under mild regularity assumptions (twice-differentiable cost, compact domains, qualified subproblem solvability), the block-coordinate minimization plus dual update scheme converges to a first-order critical point, with primal and dual residuals vanishing in the limit.

The stochastic Riemannian MARS-ADMM [2512.22750] proves iteration complexity $\tilde{\mathcal{O}}(\epsilon^{-3})$ to reach $\epsilon$-KKT points, matching deterministic ADMM rates and improving on previous stochastic Riemannian operator-splitting bounds ($\mathcal{O}(\epsilon^{-3.5})$). Convergence relies on balancing rates for penalty, primal, dual, and momentum components and employing a merit function that couples expected augmented Lagrangian decrease with control of estimator variance.

## 5. Decentralization and Implementation Aspects

The first MARS-ADMM [2111.07016] is implemented in C++11 on Linux, with all per-segment slack and plane updates naturally parallelizable. Parameter choices are uniform, with fixed penalty ($\varrho=0.1$), log-barrier weight ($\gamma=10$), and explicit velocity and acceleration bounds for UAVs and robot arms. Termination is set via $\|\nabla\mathcal{L}\|_\infty$ thresholds, and initialization requires strictly feasible paths (e.g., RRT-connect).

Stochastic Riemannian MARS-ADMM [2512.22750] avoids double-loop/batch restarts, requiring only a constant number of stochastic gradient evaluations per iteration. Penalty parameters grow slowly with iterations ($\rho_k = c_\rho k^{1/3}$), preventing numerical instability. While theory admits constant-size batches, moderate batch sizes (50-200 samples) offer practical stability.

## 6. Quantitative Performance and Empirical Comparison

Empirical evaluations for multi-robot MARS-ADMM [2111.07016] yield:
- **Single UAV trajectory planning**: Wall-times 34 ms–18.9 s, compared to primal IPM (1,000 ms–420,000 ms), indicating 10–20× speedup; trajectory cost within 0.5–1.5% of primal IPM, often shorter than Safe-Corridor.
- **4-UAV multi-group swaps**: 100% success rate vs. 53% for soft-penalty approaches; trajectory length/time 77.8 m/11.3 s vs. 80.1 m/20.6 s.
- **7-DOF KUKA arms swap**: RRT path refined from 6.38 m to 1.63 m in 16 s; convergence in ~150 iterations, sub-second per iteration.

Ablation studies show removing stiffness decoupling slows runtime by 3–5× without significant quality change; log-barrier weight $\gamma$ variations ($1$, $10$, $100$) have negligible impact on trajectory quality, but may affect iteration count.

MARS-ADMM in the Riemannian stochastic optimization setting [2512.22750] achieves near-optimal complexity, closing the theoretical gap between stochastic and deterministic operator-splitting methods.

## 7. Comparative Analysis and Implications

A summary of performance and theoretical features is presented below.

| Variant                | Domain            | Complexity Bound      | Decentralization      |
|------------------------|-------------------|----------------------|----------------------|
| Multi-robot MARS-ADMM  | Trajectory Opt.   | Empirical order-of-mag. speedup | Full parallelization (per segment/plane) |
| Stochastic Riemannian MARS-ADMM | Composite nonsmooth optimization | $\tilde{\mathcal{O}}(\epsilon^{-3})$ | Single loop, constant samples |

Both approaches demonstrate that structured nonconvex constraints and composite objectives can be split and solved efficiently via ADMM when carefully coupled with stiffness decoupling and, in the stochastic case, momentum-based variance reduction. This suggests a broad applicability of MARS-ADMM in settings where collision safety, robustness, and decentralized execution are critical, and where optimization problems reside on manifolds or require stochastic or composite objective treatment.

A plausible implication is that the methodological principles from the original multi-robot setting—stiffness decoupling, per-block slack variables, parallelism—are extensible to more abstract or general manifolds, as evidenced by the single-loop stochastic Riemannian MARS-ADMM formulation. The closing of the complexity gap to deterministic methods indicates potential for real-time and large-scale nonsmooth optimization applications under mild regularity assumptions.

**References**:  
"Robust Multi-Robot Trajectory Optimization Using Alternating Direction Method of Multiplier" [2111.07016]  
"A Single-loop Stochastic Riemannian ADMM for Nonsmooth Optimization" [2512.22750]

Source: https://www.emergentmind.com/topics/mars-admm