ADMM-DTN: Distributed Trajectory Negotiation

• ADMM-DTN is a distributed trajectory planning framework where agents negotiate local trajectories via consensus ADMM to enforce global safety and coordination.
• It decomposes complex trajectory optimization into locally solvable subproblems with iterative dual and primal updates, enhancing computational efficiency in multi-agent systems.
• Robust collision avoidance is achieved through techniques like Buffered Voronoi Cells and feasibility-preserving methods, balancing optimality with safety.

ADMM-based Distributed Trajectory Negotiation (ADMM-DTN) is a family of distributed trajectory planning and coordination methods in which multiple agents, robots, or vehicles optimize local trajectories while iteratively enforcing agreement on shared plans, timing variables, collision margins, or other coupled quantities through the Alternating Direction Method of Multipliers. In this literature, each agent typically keeps local copies of global or neighbor-relevant variables, solves a local constrained subproblem, exchanges updated plans with neighbors or a coordination layer, and updates dual variables until primal and dual residuals are sufficiently small. The term ADMM-DTN appears explicitly in distributed connectivity-maintenance and robust cooperative vehicle-coordination frameworks, while the underlying consensus-ADMM structure is developed more generally in distributed MPC and multi-robot optimization (Shetty et al., 2020, Bai et al., 29 Aug 2025, Summers et al., 2012, Chen, 2024).

1. Conceptual foundations and historical development

ADMM-DTN emerged from the convergence of distributed optimal control, consensus optimization, and multi-agent motion planning. In distributed MPC formulations, a network of dynamical subsystems is organized over a communication graph, each agent holds a local variable containing its own and neighbors’ state-input trajectories, and overlapping copies are tied together by consistency constraints. This yields a separable objective with coupling handled by augmented-Lagrangian terms and local communication, rather than a centralized solve (Summers et al., 2012).

Consensus ADMM provides the canonical negotiation mechanism. In its robotics form, each agent minimizes a local cost $g_i(\theta_i)$ subject to consensus constraints $\theta_i=\theta_j$ over graph edges, and dual variables accumulate disagreement over time. A tutorial treatment makes explicit the link from augmented Lagrangians to a neighbor-wise trajectory negotiation step in which each robot updates only from its direct neighbors’ current plans (Chen, 2024).

Early trajectory-planning variants also adopted message-passing interpretations. A factor-graph formulation for multi-agent trajectory planning represented collision-avoidance and trajectory-cost terms as local minimizers connected to equality nodes, and used an improved ADMM variant, the Three-Weight Algorithm, to exploit inactive constraints via zero weights. That work emphasized that the method is naturally parallelizable, scales well with the number of agents for several cost functionals, and can also specialize to local motion planning in velocity space (Bento et al., 2013).

A recurring theme across these formulations is that “negotiation” is not an external protocol layered over planning; it is the optimization process itself. Trajectory proposals, consensus variables, and dual multipliers jointly encode compromise between local objectives and global compatibility. A plausible implication is that ADMM-DTN is best understood not as a single algorithm, but as a design pattern for decomposing trajectory optimization into locally solvable subproblems with iterative reconciliation.

2. Core optimization structure and negotiation mechanics

A representative multi-robot formulation defines, for each robot $i$, state $x^i_t=[p^i_t,v^i_t]\in\mathbb{R}^6$, control $u^i_t=[a^i_t]\in\mathbb{R}^3$, and linear dynamics

$x^i_{t+1}=Ax^i_t+Bu^i_t.$

Over a finite horizon, each robot seeks to reach a target position, track a trajectory, avoid collisions, and respect actuator bounds (Chen, 2024).

Consensus ADMM introduces a global variable such as

$\theta=[x^1,\dots,x^N,u^1,\dots,u^N],$

with each agent $i$ maintaining a local copy $\theta^i$. A standard distributed form is

$$\begin{array}{ll} \underset{\theta^1,\dots,\theta^N}{\text{minimize}} & \sum_{i=1}^N L^i(\theta^i) \ \text{subject to} & \theta^i=\theta^j,\ \forall j\in\mathcal{N}^i, \ & g^i(\theta^i)\le 0, \ & h^i(\theta^i)=0. \end{array}$$

The local primal step then takes the form

$\theta_i=\theta_j$0

followed by the dual update

$\theta_i=\theta_j$1

with termination when primal and dual residuals fall below thresholds (Chen, 2024).

An equivalent formulation used in distributed MPC introduces a consensus variable $\theta_i=\theta_j$2, local variables $\theta_i=\theta_j$3, and componentwise averaging in the consensus step,

$\theta_i=\theta_j$4

In that setting, the $\theta_i=\theta_j$5-updates are local convex subproblems, the $\theta_i=\theta_j$6-update acts as an averaging-based reconciliation of “opinions,” and all steps except the $\theta_i=\theta_j$7-update can be done fully in parallel (Summers et al., 2012).

In receding-horizon implementations, this negotiation is embedded inside MPC. At each time step, agents solve over a horizon $\theta_i=\theta_j$8 using an inner ADMM loop, apply only the first control, shift the horizon, and repeat. This architecture appears in both multi-robot waypoint navigation and distributed MPC consensus, and it is central to real-time deployment because it permits truncation of the negotiation rounds when deadlines are tight (Chen, 2024, Summers et al., 2012).

3. Collision avoidance, convexification, and safety guarantees

The principal technical difficulty in ADMM-DTN is that collision avoidance is usually non-convex. A conventional pairwise safety constraint is

$\theta_i=\theta_j$9

which directly couples agents and breaks the convexity assumptions under which consensus ADMM has guaranteed convergence (Chen, 2024).

One influential remedy is convexification by Buffered Voronoi Cells (BVCs). For agent $i$0, the buffered cell is defined in 2D by

$i$1

where $i$2. Enforcing $i$3 at all time steps converts collision avoidance into linear local inequalities. In a waypoint-navigation study, the convex C-ADMM algorithm with BVCs required 1000 fewer iterations to achieve convergence in a multi-robot waypoint navigation scenario, and the reported average iterations were 701 versus 1713 for 3 agents and 1673 versus 4984 for 5 agents. The convex variant satisfied safety at all times and exhibited primal residuals that converge linearly to zero, whereas the non-convex baseline led to sub-optimal solutions and violation of safety constraints in trajectory generation (Chen, 2024).

A different approach is to preserve feasibility rather than convexify the geometry. A robust multi-robot trajectory-optimization method starts from a collision-free initial trajectory, introduces slack variables, and combines ADMM with log-barrier terms and CCD-based line search. The reported consequence is guaranteed collision avoidance and homotopy preservation throughout optimization, while being orders of magnitude faster than a primal interior point method; the paper reports 10–100× speedup in multi-arm and multi-UAV settings (Ni et al., 2021).

Uncertainty-aware vehicle coordination pushes safety handling further by controlling trajectory distributions instead of deterministic trajectories. In that formulation, collision chance constraints are reformulated into deterministic LMIs, the safety distance $i$4 is adapted online through an $i$5-regularization term, and dual variables are introduced for collision constraints between neighboring vehicles. The resulting fully parallel ADMM-DTN framework reports collision-rate reductions of up to 40.79\% in various scenarios, with an interactive attention mechanism reducing computational demand by 14.1\% (Bai et al., 29 Aug 2025).

These formulations illustrate a central divide in ADMM-DTN. Some methods obtain robust behavior by making the coupled safety constraints convex; others retain non-convex structure but enforce feasibility through barriers, projections, or probabilistic reformulations. This suggests that the performance of ADMM-DTN depends at least as much on the chosen safety representation as on the ADMM iterations themselves.

4. Architectural variants and representative formulations

The ADMM-DTN literature spans synchronous, asynchronous, convexified, multi-block, and uncertainty-aware architectures. The common thread is decomposition of a coupled trajectory problem into smaller subproblems joined by consensus or consistency constraints.

Setting	ADMM structure	Reported property
Distributed MPC consensus	Local copies plus averaging-based consensus variable $i$6	A few tens of ADMM iterations yield near-centralized performance
Convex multi-robot planning	Consensus ADMM with BVC collision convexification	1000 fewer iterations than a non-convex baseline
Connectivity maintenance under uncertainty	Local copies of joint nominal trajectories plus connectivity cost	1000 Monte Carlo rollouts statistically validate connectivity maintenance
Robust CAV coordination under uncertainty	Fully parallel ADMM-DTN with Jacobi-style negotiation	Up to 40.79\% lower collision rate
Communication-constrained UAV swarms	PDDP local solves with async-ADMM, partial barrier, bounded delay	Near synchronous performance for $i$7

Beyond standard consensus ADMM, several papers modify the decomposition itself. A multi-block ADMM framework for legged locomotion splits the problem into centroidal, whole-body, and projection blocks, so that rigid-body dynamics remain in unconstrained DDP/iLQR-style subproblems while box, torque, and cone constraints are handled by projections. The Stage-wise Accelerated ADMM scheme combines over-relaxation and a varying-penalty mechanism; on flat terrain, cost convergence occurred after about 15 ADMM iterations per walking step, and on rough terrain feasible robust locomotions were produced after about 30 iterations (Zhou et al., 2020).

For nonlinear cooperative driving, one line of work reformulates each convexified iLQR iteration so that the dual has a consensus structure. A dual consensus ADMM then solves the decentralized subproblem, after which a dynamically feasible nominal-trajectory update is recovered through an LQR-style forward pass and line search. In reported traffic scenarios, the multi-process implementation was 1.5x faster than centralized iLQR in a 3-vehicle T-junction and 4.7x faster in a 12-vehicle intersection, while SQP failed to converge in the larger case (Huang et al., 2023).

Asynchronous operation becomes necessary when communication is unreliable. A communication-aware two-tier architecture for UAV swarms uses PDDP for local planning and async-ADMM for coordination, with a partial barrier that allows master updates after receiving only $i$8 worker messages, and a bounded-delay condition that incorporates each worker within at most $i$9 master iterations. Simulations report that for communication success probability $x^i_t=[p^i_t,v^i_t]\in\mathbb{R}^6$0, the method achieves results near those of idealized synchronous distributed optimization in both solution quality and time efficiency (Yu et al., 19 Nov 2025).

5. Applications and empirical behavior

ADMM-DTN has been applied across multi-robot waypoint navigation, distributed MPC consensus, non-myopic target tracking, connectivity maintenance, legged locomotion, connected autonomous vehicles, and route-constrained transportation.

In distributed MPC consensus for double integrators, the practical picture is that moderate negotiation depth is often enough. With 5 agents, 1 ADMM iteration per step was about 80\% worse than centralized, 2 iterations per step were within 1.5\% of centralized performance, and 10+ iterations were within 0.5\% on average. Using code-generated local solvers, each local subproblem could be solved in under 2 milliseconds, and 30 ADMM iterations per step resulted in total time under 60 milliseconds (Summers et al., 2012).

In multi-robot waypoint navigation, the convex-versus-non-convex comparison is especially revealing because both methods eventually reach goal positions, but the non-convex approach may take less safe or less optimal paths. Visual inspections in that study showed that convex C-ADMM produces smoother, safer, and more consistent agent behaviors, reinforcing the quantitative residual and iteration-count differences (Chen, 2024).

In non-myopic path planning for multi-target tracking, the decomposition is by target rather than by robot. Each target induces a local planning subproblem, and ADMM consensus automatically induces the high-level task assignments that heuristic methods otherwise impose separately. The paper describes this as detouring the task-assignment problem and letting consensus induce high-level decisions automatically among the targets; modified receding-horizon control and edge-cutting then improve real-time operation (Park et al., 2018).

Connectivity-maintenance planning introduces a different collective objective: maintaining the algebraic connectivity of a weighted graph above a threshold under motion and sensing uncertainty. The planner models belief states via Kalman filtering, uses a conservative distance metric in the graph weights, and penalizes loss of connectivity through

$x^i_t=[p^i_t,v^i_t]\in\mathbb{R}^6$1

To reduce computational load, the Hessian of the connectivity cost is approximated by a rank-1 outer product, and the method is statistically validated with 1000 Monte Carlo rollouts (Shetty et al., 2020).

Route-constrained multi-agent transportation replaces spatial rerouting by timing negotiation. In an inexact-projection ADMM scheme, waypoint passage times are optimized while preserving each agent’s waypoint order and nominal route assignment, and pairwise safety is encoded through distance-based penalties on a dense temporal grid. In bottleneck cases, the reported total mission times for a representative hierarchical baseline versus the proposed method were 172.5 s versus 152.0 s at 1\% occupancy, 221.3 s versus 171.4 s at 5\%, and infeasible versus 187.2 s at 10\% (Lee et al., 23 Mar 2026).

6. Limitations, misconceptions, and evolving directions

A persistent misconception is that ADMM itself guarantees reliable convergence in robotic trajectory negotiation. The literature is more specific: C-ADMM is well known for guaranteed convergence in convex optimization problems, while non-convex robotics formulations often lack theoretical guarantees and may converge to stationary points, sub-optimal solutions, or unsafe trajectories. The convex-versus-non-convex multi-robot study makes this contrast explicit, and the robotics tutorial also distinguishes the convex and non-convex cases (Chen, 2024, Chen, 2024).

Another misconception is that distributed optimization automatically removes communication bottlenecks. Online distributed ADMM shows that performance depends on the underlying network topology and condition measures associated with the linear constraints, with social regret bounded sublinearly as $x^i_t=[p^i_t,v^i_t]\in\mathbb{R}^6$2 under standard assumptions. This formalizes the role of graph connectivity in consensus quality and speed (Hosseini et al., 2014).

Penalty and relaxation parameters are also more consequential than early presentations of ADMM-DTN sometimes imply. Stage-wise varying-penalty schedules, spectral-gradient penalty tuning, and learned agent-wise penalties all appear as responses to slow or poorly balanced residual contraction. A spatial-temporal UAV framework reports that spectral adaptive penalty updates reduce the number of ADMM iterations by 20–50\% and total solve time by similar margins, while a differentiable ADMM-DDP meta-learning framework reports up to $x^i_t=[p^i_t,v^i_t]\in\mathbb{R}^6$3 faster gradient computation than state-of-the-art methods by reusing DDP components and solving an auxiliary ADMM-coordinated gradient problem (Zheng et al., 20 Oct 2025, Wang et al., 1 Sep 2025).

Second-order and acceleration mechanisms are likewise becoming central. Distributed Newton-ADMM establishes global linear convergence for strongly convex smooth-plus-nonsmooth consensus optimization by approximating the Hessian inverse with a truncated Taylor expansion, requiring exactly $x^i_t=[p^i_t,v^i_t]\in\mathbb{R}^6$4 distributed communication rounds per Newton step. This is not a trajectory-planning result by itself, but it provides a technically relevant template for faster distributed consensus updates in optimization layers that resemble ADMM-DTN (Li et al., 2021).

The present trajectory of the field points toward four converging directions: stronger safety handling under uncertainty, tighter integration with local optimal-control solvers such as DDP and iLQR, robustness to asynchronous and unreliable communication, and systematic reduction of hand-tuned hyperparameters through adaptive or learned coordination mechanisms. This suggests that ADMM-DTN is evolving from a consensus solver used inside trajectory planners into a broader systems framework for distributed, uncertainty-aware, and resource-constrained multi-agent coordination.