Contract MPC Enables Anytime Robot Traffic Coordination

Multi-robot traffic systems break many distributed MPC assumptions. The communication graph changes as vehicles enter, leave, or move out of range, while collision avoidance still has to hold without a central scheduler. This paper addresses that setting with an anytime plug-and-play controller: agents coordinate only with distance-defined neighbors, exchange predicted trajectories, and construct local contracts that preserve safety as the network changes.

The authors’ main claim is stronger than standard distributed collision avoidance. They argue that agents can plug in or out at arbitrary times, without request-based admission, pre-negotiated schedules, or centralized coordination, while keeping recursive feasibility and collision-free operation. Figure 1 gives the experimental setting: small-scale autonomous race cars running on a figure-eight track, in simulation and hardware, with up to eight vehicles.

Core Contribution

The contribution is a contract-based distributed MPC scheme for nonlinear multi-agent systems with time-varying local communication. Prior plug-and-play MPC often assumes fixed participants, request-based joining, centralized checks, or a communication topology that changes only within predefined limits. Here, the contracts are derived from predicted trajectories and neighbor relationships, so the controller can update its safety obligations as nearby agents appear or disappear.

The paper’s comparison table is useful because it states the intended trade-off explicitly. Against 11 prior approaches listed in Table I, the proposed method is the only row marked as distributed, compatible with nonlinear dynamics, anytime plug-and-play, safety-guaranteed, non-cooperative, non-iterative, and non-sequential. That is an author-organized comparison, not a benchmark, but it clarifies the niche: the method sacrifices some flexibility through conservative envelopes in order to remove centralized coordination and negotiation loops.

Technical Approach

Each agent solves a local MPC problem with collision-avoidance constraints built from two geometric objects: cells and safety envelopes. The cells partition the predicted multi-agent motion over the MPC horizon using a Voronoi-type construction, while the envelopes bound how far an agent may move from its communicated prediction between updates. The result is a contract: if every agent remains inside its assigned admissible region, pairwise separation is preserved.

Figure 3 shows how those cells change along the horizon as the predicted positions move. The important detail is that the partition is not static. It is recomputed from the current predicted trajectories, which is what lets the method follow a changing local neighborhood rather than a fixed graph.

The formal part of the paper establishes recursive feasibility and collision avoidance under the stated assumptions. The appendix proofs show how disjoint inflated cells imply a separation margin of at least $\epsilon$, and how the safety-envelope construction bounds accumulated deviation along the horizon. In plain terms, the controller does not need every vehicle to agree on a global plan; it needs local predictions and contracts that leave enough geometric slack for the next MPC update.

Results and Analysis

The empirical validation uses autonomous multi-agent driving, including simulation and hardware experiments. The paper reports collision-free behavior in high-speed, dynamic scenes with agents moving in opposite directions. The most informative qualitative case is the figure-eight scenario in Figure 7: one vehicle moves away from the centerline to avoid oncoming traffic, and another slows and diverts at a busy intersection until the crossing agents clear the conflict region.

The evidence supports the safety and feasibility claims more directly than it supports performance claims. The formal proofs address the core guarantee, and the hardware demonstrations show that the controller can run on real small-scale vehicles rather than only in a symbolic example. The experiments also stress the communication topology in a way that matches the paper’s motivation: nearby agents change over time, and the controller reacts through local contracts rather than a central traffic manager.

What is missing is a quantitative performance study. The available results do not report collision-rate comparisons, solve times, tracking error distributions, throughput, or ablations against request-based plug-and-play MPC. That makes the paper convincing as a control framework and proof-backed demonstration, but less complete as an empirical systems paper. The key practical question is not whether the contracts can enforce safety under the assumptions; the paper gives a clear argument for that. It is how much conservatism the safety envelopes introduce as density, speed, or communication delay grows.

Limitations

The authors identify the main weakness themselves: safety envelopes can limit per-step progress, creating conservative behavior even when more aggressive motion would be physically safe. The framework also still requires neighbor-to-neighbor information exchange, so it does not solve communication-free distributed control. Finally, the comparison to prior work is mostly assumption-level rather than experimental; the method’s advantage is clearest in scenarios where anytime joining and leaving matter more than optimality or traffic throughput.