Preemptive Holistic Collaborative System

• PHCS is a comprehensive multi-agent coordination framework that enables autonomous agents to share intentions and preemptively resolve conflicts for collaborative trajectory planning.
• It employs receding-horizon control, hybrid automata, and conflict detection to optimize both individual and system-level objectives in real time.
• The framework scales across diverse domains—from road transportation to manufacturing—demonstrating significant efficiency gains and enhanced safety in simulation studies.

The Preemptive Holistic Collaborative System (PHCS) is a multi-agent coordination framework that enables independent entities to collaboratively plan and execute spatio-temporal trajectories by sharing intentions and preemptively resolving conflicts. The PHCS architecture generalizes across domains such as road transportation, manufacturing, supply chains, and multi-robot systems, facilitating joint optimization of both individual and system-level objectives while guaranteeing safety and scalability. In close connection with model predictive control (MPC) and hybrid automaton formalism, PHCS synchronizes agent plans over receding horizons, employs conflict detection with preemptive resolution, and can be hierarchically decomposed to manage complexity in large-scale embodied multi-agent systems (Li et al., 2024, Peng, 6 Jan 2026).

1. Mathematical Formulation and Principles

PHCS models a set $E = \{1, 2, ..., N\}$ of agents (entities) with individual dynamics

$\dot x_i(t) = f_i(x_i(t), u_i(t))$

where $x_i(t)$ encodes the state (e.g., position, velocity), $u_i(t)\in U_i$ is the control input, and $\theta_i$ is a static intent comprising, for instance, target lane or desired arrival window. Each agent seeks to minimize a private cost functional

$$J_i[x_i(\cdot), u_i(\cdot)] = \int_0^H \ell_i(x_i(t), u_i(t))\, dt + \phi_i(x_i(H))$$

subject to constraints $x_i(t) \in X_i$, $u_i(t) \in U_i$, and collision avoidance $\|p_i(t) - p_j(t)\| \geq d_{\mathrm{safe}}$ for $i \neq j$.

PHCS employs periodic information sharing: at each planning epoch $\dot x_i(t) = f_i(x_i(t), u_i(t))$0, every agent $\dot x_i(t) = f_i(x_i(t), u_i(t))$1 broadcasts its state $\dot x_i(t) = f_i(x_i(t), u_i(t))$2, intended maneuvers $\dot x_i(t) = f_i(x_i(t), u_i(t))$3, and most recently planned trajectory $\dot x_i(t) = f_i(x_i(t), u_i(t))$4 to a local coordinator (e.g., a Road Section Management Unit, RSMU). The coordinator aggregates all submissions, forecasts pairwise conflicts, and—if necessary—solves a joint optimization problem:

$\dot x_i(t) = f_i(x_i(t), u_i(t))$5

subject to system dynamics and collision constraints, with $\dot x_i(t) = f_i(x_i(t), u_i(t))$6 a penalty function such as $\dot x_i(t) = f_i(x_i(t), u_i(t))$7 (Li et al., 2024). This ensures alignment of individual goals with holistic safety and efficiency.

2. Hybrid Automaton and Three-Stage Receding Horizon Protocol

The core timing and update cycle of PHCS is modeled via a hybrid automaton for the coordinator:

$\dot x_i(t) = f_i(x_i(t), u_i(t))$8

with discrete modes $\dot x_i(t) = f_i(x_i(t), u_i(t))$9, a continuous timer $x_i(t)$0, and guards/enforcements to prevent Zeno behavior (infinite switching), requiring $x_i(t)$1 (Peng, 6 Jan 2026).

Each control cycle subdivides the look-ahead horizon $x_i(t)$2 into:

• Frozen Window ($x_i(t)$3): Agents execute an immutable trajectory prefix.
• Planning Window ($x_i(t)$4): Coordinators solve for new plans.
• Look-ahead Window ($x_i(t)$5): Coordinators check for imminent future conflicts ("preemptive" action).

Explicit padding ($x_i(t)$6) is used during $x_i(t)$7 to avoid race conditions between plan dissemination and new intent updates, ensuring temporal separation and robust handoff.

Recommended ratios are:

$x_i(t)$8

3. Information-Sharing Architecture and Shadow Agent Protocol

The PHCS architecture distinguishes between physical agents and coordinators. Each agent hosts a Vehicle Intelligent Unit (VIU) for intention encoding and communication. Coordinators (such as RSMUs) are deployed across subspaces, maintaining local and boundary-aware optimization. Communication is facilitated by DSRC (IEEE 802.11p) for local V2I/V2V and cellular channels for global coordination.

For hierarchical scalability, the workspace $x_i(t)$9 is partitioned into subspaces, each managed by a logical coordinator $u_i(t)\in U_i$0. When agents cross boundaries, the "Shadow Agent Protocol" is invoked: each agent shares its spatiotemporal tube $u_i(t)\in U_i$1 with neighboring coordinators, which instantiate shadow agents and enforce state separation coupled by an Input-to-State Stability (ISS) penalty. Discrepancies $u_i(t)\in U_i$2 between agent and shadow are penalized via a Lyapunov function, ensuring bounded disagreement under communication, modeling, or solver errors (Peng, 6 Jan 2026).

4. Conflict Detection, Preemptive Resolution, and Computational Complexity

Conflict detection operates with $u_i(t)\in U_i$3 complexity per cycle: for each agent pair $u_i(t)\in U_i$4, the coordinator predicts respective trajectories and if $u_i(t)\in U_i$5 within the planning horizon, a potential conflict is flagged.

Upon conflict detection, the coordinator executes the following sequence:

1. Snapshot: Gather all current states, intents, and planned trajectories;
2. Conflict Detection: Identify conflicting pairs;
3. Joint Optimization: Solve the above holistic program (typically via QP or sequential QP methods) over all affected agents;
4. Dissemination: Return updated trajectory plans $u_i(t)\in U_i$6 (Li et al., 2024).

For up to a few dozen agents and horizon $u_i(t)\in U_i$7, modern QP solvers achieve solution times within $u_i(t)\in U_i$825 ms per cycle, satisfying real-time operation requirements. In the hierarchical extension, local optimization problems are of size $u_i(t)\in U_i$9 and the global computational load becomes $\theta_i$0 per subspace, manageable by subdivision (Peng, 6 Jan 2026).

5. Application to Road Transportation: PHCRTS

The Preemptive Holistic Collaborative Road Transportation System (PHCRTS) is a direct instantiation of PHCS. Each vehicle is represented by a reduced kinematic bicycle model:

$\theta_i$1

with discrete lane index $\theta_i$2. Intents $\theta_i$3 specify target lane, desired merge point, and allowable merge time. Trajectories are parameterized via polynomials or B-splines.

The specific collaborative optimization in highway merging minimizes:

$\theta_i$4

subject to vehicle dynamics, lane-keeping, and safe-distance constraints (Li et al., 2024).

6. Empirical Performance and Theoretical Guarantees

In two-lane highway merging simulations (road: 500 m + 200 m ramp, $\theta_i$5 m, $\theta_i$6 m/s, $\theta_i$7 m/s$\theta_i$8, $\theta_i$9 s, repeat 20 trials), PHCRTS demonstrated:

Metric	Baseline	PHCRTS	Relative Change
Avg. time delay	30.5 s	3.2 s	$$J_i[x_i(\cdot), u_i(\cdot)] = \int_0^H \ell_i(x_i(t), u_i(t))\, dt + \phi_i(x_i(H))$$0 ≈90%
Throughput (veh/h)	1000	4000	$$J_i[x_i(\cdot), u_i(\cdot)] = \int_0^H \ell_i(x_i(t), u_i(t))\, dt + \phi_i(x_i(H))$$1 300%
Accident count/hr	12	0	Eliminated

Under PHCRTS, agents (vehicles) merge without full stops, delays are minimized especially in moderate traffic, and solver times enable real-time closed-loop operation. The framework's safety is theoretically underpinned by tube-inflated separation, dwell-time enforced Zeno-freeness, recursive feasibility over receding horizons, and probabilistic safety under communication dropout (risk $$J_i[x_i(\cdot), u_i(\cdot)] = \int_0^H \ell_i(x_i(t), u_i(t))\, dt + \phi_i(x_i(H))$$2 via blackout-resilient planning) (Li et al., 2024, Peng, 6 Jan 2026).

7. Scalability, Extensions, and Open Challenges

PHCS generalizes to any domain where independent agents interact in a shared physical or logical space—examples include port vessel scheduling, industrial robotics, and smart manufacturing lines. The introduction of hierarchical decomposition and the Shadow Agent protocol provides a scalable path, capping computational growth and removing single points of failure.

Key open challenges include:

• Handling malicious or uncooperative agents within the information-sharing loop;
• Further reducing computational requirements as $$J_i[x_i(\cdot), u_i(\cdot)] = \int_0^H \ell_i(x_i(t), u_i(t))\, dt + \phi_i(x_i(H))$$3 grows, potentially through advanced hierarchical or distributed decompositions;
• Integrating human-in-the-loop adaptation and override;
• Ensuring robust operation under lossy communication and asynchronous updates.

Formally, PHCS/Prollect synthesizes hybrid automaton timing, receding-horizon MPC, preemptive look-ahead, and ISS-based shadow handover to deliver collision-free, stable, scalable, and provably robust multi-agent coordination (Peng, 6 Jan 2026).