TurboADMM: Efficient Multi-Agent QP Solver

• TurboADMM is a specialized solver for multi-agent trajectory optimization that employs ADMM decomposition to preserve per-agent temporal structure while managing dense coupling constraints.
• It leverages Riccati warmstart for high-quality initialization and qpOASES hotstart to reuse KKT factorizations, ensuring rapid convergence even as agent counts grow.
• Empirical evaluations demonstrate near-linear scaling with agents and significant runtime improvements over general-purpose solvers like OSQP and MOSEK.

TurboADMM most specifically denotes a specialized quadratic-program solver for centralized multi-agent trajectory optimization introduced in 2026, designed for settings with dense inter-agent coupling, linear dynamics, and real-time model predictive control requirements (Chen, 14 Jan 2026). In a broader and less standardized usage, the term has also functioned as shorthand for ADMM variants that seek substantially faster practical convergence through acceleration, adaptive penalties, preconditioning, variance reduction, or learned iteration policies. This broader usage is not uniform across the literature. The explicit named method, however, is a single-machine, shared-memory solver that co-designs ADMM decomposition, Riccati warmstart, and parametric QP hotstart so that per-agent temporal structure is preserved while dense coupling constraints are handled through consensus variables (Chen, 14 Jan 2026).

1. Terminology, scope, and historical placement

The named method "TurboADMM" appears as "TurboADMM: A Structure-Exploiting Parallel Solver for Multi-Agent Trajectory Optimization" and targets large coupled QPs arising from centralized multi-agent trajectory optimization with dense interaction networks (Chen, 14 Jan 2026). Its stated motivation is that general-purpose QP solvers such as OSQP and MOSEK treat the problem monolithically and encounter scalability difficulties with agent count, while structure-exploiting solvers such as HPIPM leverage temporal structure through Riccati recursion but can be vulnerable to dense coupling constraints (Chen, 14 Jan 2026).

Outside that explicit usage, the literature reflected in related ADMM papers uses "TurboADMM" more informally. Several works characterize "turbo" ADMM as ADMM augmented by acceleration, adaptive penalty selection, over-relaxation, variance reduction, or adaptive proximal geometry rather than as a single canonical algorithm. "Fast Stochastic Alternating Direction Method of Multipliers" presents an incremental full-gradient stochastic ADMM that improves the convergence rate on convex problems from $O(1/\sqrt{T})$ to $O(1/T)$, and its tutorial-style reconstruction explicitly interprets this batch-like stochastic convergence as a "turbo" effect (Zhong et al., 2013). "Adaptive Stochastic Alternating Direction Method of Multipliers" instead improves stochastic ADMM by changing the proximal geometry adaptively through second-order metrics $H_t$, and its detailed discussion places it in the same family as "TurboADMM"-style methods while emphasizing adaptive metric selection rather than momentum or Nesterov acceleration (Zhao et al., 2013).

This suggests that the term has two distinct meanings in current usage. First, it is the proper name of a solver for multi-agent MPC-style QPs (Chen, 14 Jan 2026). Second, it is a broader editorial label for ADMM variants engineered for markedly faster convergence, often by exploiting additional structure beyond the baseline two-block splitting. A common misconception is therefore that TurboADMM refers only to momentum-based ADMM. The surrounding literature does not support that restriction: acceleration has also been pursued through spectral penalty adaptation (Xu et al., 2016), Anderson acceleration (Zhang et al., 2019), stochastic variance reduction (Zhong et al., 2013), adaptive subgradient metrics (Zhao et al., 2013), stochastic dual coordinate updates (1311.0622), and learned step sizes and communication weights in distributed ADMM (Doerks et al., 5 Sep 2025).

2. Problem class and mathematical formulation

The explicit TurboADMM solver addresses centralized multi-agent trajectory optimization problems in which there are $N$ agents, each with linear dynamics over a planning horizon $T$, local state and input constraints, and dense all-pairs collision constraints at every time step (Chen, 14 Jan 2026). For agent $i$, the dynamics are

$$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$

with box constraints

$$x_{i,t} \in \mathcal{X}_i,\quad u_{i,t} \in \mathcal{U}_i.$$

The cost is quadratic tracking plus quadratic control effort: $$\begin{aligned} \min_{\{x_{i,t},u_{i,t}\}} \ &\sum_{i=1}^N \left[ \sum_{t=0}^{T-1} \left( \|x_{i,t}-x_{i,\text{ref},t}\|_{Q_i}^2 + \|u_{i,t}\|_{R_i}^2 \right) \right. \ &\left. + \|x_{i,T}-x_{i,\text{ref},T}\|_{Q_{i,T}}^2 \right]. \end{aligned}$$ Collision avoidance is imposed after linearization through constraints of the form

$$\|p_{i,t} - p_{j,t}\| \ge d_{\text{safe}},\quad \forall i<j,\ t,$$

where $O(1/T)$0 extracts position (Chen, 14 Jan 2026).

The key structural asymmetry is that dynamics are block-tridiagonal in time for each agent, whereas collision constraints are dense across agents at each time step. As $O(1/T)$1 grows, the number of state and control variables grows linearly with $O(1/T)$2, but the number of collision constraints grows quadratically with $O(1/T)$3, $O(1/T)$4 (Chen, 14 Jan 2026). This destroys simple global sparsity patterns in a monolithic QP formulation.

TurboADMM addresses this by introducing consensus variables $O(1/T)$5 for each pair $O(1/T)$6 and time $O(1/T)$7, so that dense coupling is moved out of the local agent constraints and into consensus and dual updates (Chen, 14 Jan 2026). The resulting ADMM iteration has three steps: $O(1/T)$8

$O(1/T)$9

$H_t$0

with $H_t$1, fixed to $H_t$2 in the reported experiments (Chen, 14 Jan 2026).

The local per-agent QP retains the temporal structure: $H_t$3 where $H_t$4 is block-tridiagonal and the ADMM penalty modifies only the local cost matrices through additive terms such as $H_t$5 (Chen, 14 Jan 2026). This preservation of local temporal structure is central to the solver design.

3. Co-designed algorithmic components

TurboADMM is defined by the co-design of three complementary components rather than by ADMM splitting alone (Chen, 14 Jan 2026).

Component	Function	Structural role
ADMM decomposition	Creates per-agent subproblems solvable in parallel	Preserves block-tridiagonal local structure
Riccati warmstart	Provides high-quality primal-dual initialization for each agent's QP	Exploits temporal structure
qpOASES hotstart	Reuses similar KKT factorizations across ADMM iterations	Exploits iteration similarity

The ADMM decomposition removes inter-agent coupling from the local constraint matrices. For agent $H_t$6, all interactions with other agents appear only as quadratic penalties on $H_t$7 relative to $H_t$8 and $H_t$9, so each local QP can be solved independently and in parallel on a shared-memory machine using OpenMP (Chen, 14 Jan 2026).

The Riccati warmstart is applied in the first ADMM iteration for each agent. TurboADMM temporarily ignores inequality constraints, solves the resulting unconstrained affine LQR problem via Riccati recursion, and uses the resulting trajectories and dynamics multipliers as a primal-dual initialization for qpOASES (Chen, 14 Jan 2026). The backward pass uses recursions such as

$N$0

with $N$1, followed by a forward rollout

$N$2

and dynamics dual warmstart

$N$3

(Chen, 14 Jan 2026).

The qpOASES hotstart operates across ADMM iterations. Because the local QP structure remains unchanged while only linear terms and right-hand sides vary slightly through updates in $N$4 and $N$5, qpOASES can reuse KKT factorizations and the previous active set (Chen, 14 Jan 2026). The paper distinguishes terminology carefully: "warmstart" means using Riccati results as auxiliary initialization for a single QP solve, while "hotstart" means reusing qpOASES internal factorization and active-set information across QP solve iterations (Chen, 14 Jan 2026).

A plausible implication is that TurboADMM is best understood not as a single acceleration trick, but as a layered architecture. ADMM supplies decomposition, Riccati recursion supplies a structure-exploiting initial point, and qpOASES hotstart amortizes repeated solves over the outer ADMM loop.

4. Computational behavior and empirical evaluation

The reported complexity model separates per-agent and global costs. For one ADMM iteration, Riccati warmstart costs

$N$6

while the qpOASES solve costs

$N$7

where $N$8 is the number of active-set iterations (Chen, 14 Jan 2026). Asymptotically the qpOASES term dominates, but the paper states that with effective warmstart and hotstart, $N$9 is empirically very small and nearly independent of $T$0 (Chen, 14 Jan 2026).

The global wall-clock cost per ADMM iteration is approximately the cost of the slowest per-agent QP because local QPs are solved in parallel (Chen, 14 Jan 2026). For the benchmark with fixed horizon $T$1, state dimension $T$2, and control dimension $T$3, the observed runtime scaling with agent count is reported as roughly $T$4, while OSQP and MOSEK scale approximately as $T$5 (Chen, 14 Jan 2026).

The experimental setup uses an Intel Core i7-155H with 22 physical cores, agent counts from 2 to 14, dense all-pairs collision avoidance with $T$6 m, reference trajectories given by straight-line paths that intersect in the middle, $T$7, and residual tolerance $T$8 (Chen, 14 Jan 2026). A representative 14-agent instance has 1736 variables, 1120 dynamics constraints, and approximately 1911 collision constraints (Chen, 14 Jan 2026).

An ablation isolates the effect of hotstart and warmstart. Total QP iterations across all agents and ADMM iterations are reported as follows (Chen, 14 Jan 2026):

Setting	2 agents	14 agents
BaseADMM	170	109,512
HotstartADMM	86	1,014
TurboADMM	4	476

At 14 agents, the corresponding CPU times are approximately BaseADMM: $T$9 s, HotstartADMM: $i$0 ms, and TurboADMM: $i$1 ms (Chen, 14 Jan 2026). The paper therefore attributes the observed speedup to the combination of hotstart and Riccati warmstart rather than to either mechanism in isolation.

Against external baselines, the reported runtimes at 14 agents are approximately TurboADMM: $i$2 ms, OSQP: $i$3 ms, and MOSEK: $i$4 ms, while HPIPM fails in the denser coupled cases (Chen, 14 Jan 2026). Trajectory quality, measured by final position error, is reported at 14 agents as TurboADMM: $i$5 m, MOSEK: $i$6 m, and OSQP: $i$7 m (Chen, 14 Jan 2026). This indicates that TurboADMM is not uniformly best on every metric, but occupies a particular regime of favorable scaling and real-time feasibility.

The HPIPM failure analysis is especially informative. The paper reports that HPIPM solves the 2-agent case, but at 4 agents, when both collision pairs are active, the interior-point method fails: the complementarity residual grows to $i$8, step sizes go to minimum, and LP relaxations confirm that the problem remains feasible and well-posed (Chen, 14 Jan 2026). This directly motivates the decomposition strategy.

5. Relation to the broader ADMM acceleration landscape

TurboADMM sits within a broad family of methods that try to preserve ADMM’s splitting advantages while reducing practical convergence time. The mechanisms used across this literature are heterogeneous.

One line of work accelerates stochastic ADMM through variance reduction. "Fast Stochastic Alternating Direction Method of Multipliers" proposes SA-ADMM and SA-IU-ADMM, which incrementally approximate the full gradient in the linearized ADMM formulation and improve the convergence rate on convex problems from $i$9 to $$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$0, matching batch ADMM without visiting all samples each iteration (Zhong et al., 2013). Another line uses adaptive second-order proximal geometry. "Adaptive Stochastic Alternating Direction Method of Multipliers" introduces quadratic Bregman divergences

$$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$1

with $$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$2 chosen from past gradients in a manner analogous to AdaGrad, yielding data-dependent regret bounds that are as good as the best fixed diagonal or full-matrix proximal choice in hindsight, up to constants (Zhao et al., 2013).

A different axis concerns automatic penalty selection. "Adaptive ADMM with Spectral Penalty Parameter Selection" interprets ADMM through dual Douglas-Rachford splitting and proposes spectral updates for the penalty parameter $$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$3, using the geometric mean of two Barzilai-Borwein-type curvature estimates together with correlation-based safeguards (Xu et al., 2016). The reported effect is fast convergence and relative insensitivity to initial stepsize and problem scaling (Xu et al., 2016).

Another class adds an external nonlinear accelerator. "Accelerating ADMM for Efficient Simulation and Optimization" represents ADMM as a fixed-point iteration and applies Anderson acceleration to either a two-variable state $$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$4 or, under separable structure, a single variable $$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$5 or $$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$6, with acceptance and restart rules based on combined or forward residuals (Zhang et al., 2019). The paper’s tutorial-style interpretation explicitly frames this as "plain ADMM + an external nonlinear accelerator" and thereby very close in spirit to a turbo layer on top of a baseline ADMM iteration (Zhang et al., 2019).

Stochastic dual-coordinate formulations offer yet another route. "Stochastic Dual Coordinate Ascent with Alternating Direction Multiplier Method" embeds SDCA-style coordinate or mini-batch updates into ADMM and proves an R-linear convergence result under mild local strong convexity assumptions, while handling structured regularizers through ADMM splitting (1311.0622). More recently, "Learning to accelerate distributed ADMM using graph neural networks" recasts distributed ADMM as a message-passing architecture and learns adaptive step sizes and communication weights, while preserving convergence by freezing learned parameters after a finite number of iterations (Doerks et al., 5 Sep 2025).

These works suggest that the 2026 TurboADMM solver should not be conflated with a single generic acceleration principle. In the broader ADMM literature, "turbo" behavior has been pursued through at least five distinct mechanisms: variance reduction, adaptive metrics, spectral penalty adaptation, nonlinear fixed-point acceleration, and learned hyperparameter policies. The named TurboADMM solver adopts none of these in their generic form; instead, it specializes to a structured QP regime where decomposition, temporal recursion, and hotstart reuse are the dominant sources of speedup (Chen, 14 Jan 2026).

6. Limitations, misconceptions, and likely extensions

The explicit TurboADMM solver assumes linear dynamics and convexified collision constraints obtained through linearization around a trajectory (Chen, 14 Jan 2026). Nonconvexity is therefore handled outside the core solver, in outer SQP-like loops that the paper does not detail (Chen, 14 Jan 2026). The method also fixes $$x_{i,t+1} = A_i x_{i,t} + B_i u_{i,t}, \quad t=0,\dots,T-1,$$7 in the reported experiments and does not study adaptive penalty scheduling, momentum, or over-relaxation, even though such mechanisms are often associated with accelerated ADMM variants (Chen, 14 Jan 2026).

Its computational model is also specific. The implementation is single-machine and shared-memory, using C++, qpOASES, OpenMP, and a central coordinator for consensus and dual updates (Chen, 14 Jan 2026). It does not address networked or asynchronous distributed computation. A plausible implication is that its near-linear empirical scaling depends on the combination of moderate horizon, fixed per-agent dimensions, and enough local parallel resources to solve the per-agent QPs concurrently.

Several misconceptions can therefore be corrected. TurboADMM is not simply ADMM with momentum; the named method does not use momentum at all (Chen, 14 Jan 2026). It is not a universally applicable black-box ADMM improvement; it is tailored to multi-agent trajectory optimization with block-tridiagonal local structure and dense pairwise coupling (Chen, 14 Jan 2026). Nor is it identical to earlier "turbo" uses in machine learning ADMM, where the focus was often stochastic variance reduction (Zhong et al., 2013), adaptive proximal metrics (Zhao et al., 2013), or spectral penalty automation (Xu et al., 2016).

The future directions stated or implied in the source material are consistent with this distinction. For the named solver, extensions include nonlinear dynamics via SQP or DDP-style outer loops, embedded deployment on resource-constrained platforms, heterogeneous agents, and improved ADMM outer-loop convergence through adaptive penalty scheduling and momentum or over-relaxation variants (Chen, 14 Jan 2026). In the broader literature, hybridization is an obvious recurring theme: adaptive metrics inside ADMM (Zhao et al., 2013), spectral penalty selection (Xu et al., 2016), Anderson-style outer acceleration (Zhang et al., 2019), or learned iteration-dependent parameters (Doerks et al., 5 Sep 2025) could in principle be combined with the structure-exploiting decomposition used by TurboADMM. This suggests that the term may continue to denote both a specific solver and a wider design philosophy: making ADMM fast by aligning the splitting method with the dominant algebraic, statistical, and hardware structure of the target problem.