Kai Cai

Kai Cai, Hideaki Ishii

We study the average consensus problem of multi-agent systems for general network topologies with unidirectional information flow. We propose two (linear) distributed algorithms, deterministic and gossip, respectively for the cases where the inter-agent communication is synchronous and asynchronous. Our contribution is that in both cases, the developed algorithms guarantee state averaging on arbitrary strongly connected digraphs; in particular, this graphical condition does not require that the network be balanced or symmetric, thereby extending many previous results in the literature. The key novelty of our approach is to augment an additional variable for each agent, called "surplus", whose function is to locally record individual state updates. For convergence analysis, we employ graph-theoretic and nonnegative matrix tools, with the eigenvalue perturbation theory playing a crucial role.

Yingying Liu, Kai Cai, Zhiwu Li

In this paper we study multi-agent discrete-event systems where the agents can be divided into several groups, and within each group the agents have similar or identical state transition structures. We employ a relabeling map to generate a "template structure" for each group, and synthesize a scalable supervisor whose state size and computational process are independent of the number of agents. This scalability allows the supervisor to remain invariant (no recomputation or reconfiguration needed) if and when there are agents removed due to failure or added for increasing productivity. The constant computational effort for synthesizing the scalable supervisor also makes our method promising for handling large-scale multi-agent systems. Moreover, based on the scalable supervisor we design scalable local controllers, one for each component agent, to establish a purely distributed control architecture. Three examples are provided to illustrate our proposed scalable supervisory synthesis and the resulting scalable supervisors as well as local controllers.

Kai Cai, Hideaki Ishii

We have recently proposed quantized gossip algorithms which solve the consensus and averaging problems on directed graphs with the least restrictive connectivity requirements. In this paper we study the convergence time of these algorithms. To this end, we investigate the shrinking time of the smallest interval that contains all states for the consensus algorithm, and the decay time of a suitable Lyapunov function for the averaging algorithm. The investigation leads us to characterizing the convergence time by the hitting time in certain special Markov chains. We simplify the structures of state transition by considering the special case of complete graphs, where every edge can be activated with an equal probability, and derive polynomial upper bounds on convergence time.

Kai Cai, Renyuan Zhang, W. M. Wonham

We study the new concept of relative coobservability in decentralized supervisory control of discrete-event systems under partial observation. This extends our previous work on relative observability from a centralized setup to a decentralized one. A fundamental concept in decentralized supervisory control is coobservability (and its several variations); this property is not, however, closed under set union, and hence there generally does not exist the supremal element. Our proposed relative coobservability, although stronger than coobservability, is algebraically well-behaved, and the supremal relatively coobservable sublanguage of a given language exists. We present an algorithm to compute this supremal sublanguage. Moreover, relative coobservability is weaker than conormality, which is also closed under set union; unlike conormality, relative coobservability imposes no constraint on disabling unobservable controllable events.

Renyuan Zhang, Kai Cai

We study supervisor localization for timed discrete-event systems under partial observation and communication delay in the Brandin-Wonham framework. First, we employ timed relative observability to synthesize a partial-observation monolithic supervisor; the control actions of this supervisor include not only disabling action of prohibitible events (as that of controllable events in the untimed case) but also "clock-preempting" action of forcible events. Accordingly we decompose the supervisor into a set of partial-observation local controllers one for each prohibitible event, as well as a set of partial-observation local preemptors one for each forcible event. We prove that these local controllers and preemptors collectively achieve the same controlled behavior as the partial-observation monolithic supervisor does. Moreover, we propose channel models for inter-agent event communication and impose bounded and unbounded delay as temporal specifications. In this formulation, there exist multiple distinct observable event sets; thus we employ timed relative coobservability to synthesize partial-observation decentralized supervisors, and then localize these supervisors into local controllers and preemptors. The above results are illustrated by a timed workcell example.

Renyuan Zhang, Kai Cai

Recently we developed supervisor localization, a top-down approach to distributed control of discrete-event systems (DES) with finite behavior. Its essence is the allocation of monolithic (global) control action among the local control strategies of individual agents. In this report, we extend supervisor localization to study the distributed control of DES with infinite behavior. Specifically, we first employ Thistle and Wonham's supervisory control theory for DES with infinite behavior to compute a safety supervisor (for safety specifications) and a liveness supervisor (for liveness specifications), and then design a suitable localization procedure to decompose the safety supervisor into a set of safety local controllers, one for each controllable event, and decompose the liveness supervisor into a set of liveness local controllers, two for each controllable event. The localization procedure for decomposing the liveness supervisor is novel; in particular, a local controller is responsible for disabling the corresponding controllable event on only part of the states of the liveness supervisor, and consequently, the derived local controller in general has states number no more than that computed by considering the disablement on all the states. Moreover, we prove that the derived local controllers achieve the same controlled behavior with the safety and liveness supervisors. We finally illustrate the result by a Small Factory example.

Kai Cai

We study a multi-agent output regulation problem, where not all agents have access to the exosystem's dynamics. We propose a distributed controller that solves the problem for linear, heterogeneous, and uncertain agent dynamics as well as time-varying directed networks. The distributed controller consists of two parts: (1) an exosystem generator that creates a local copy of the exosystem dynamics by using consensus protocols, and (2) a dynamic compensator that uses (again) consensus to approach the internal model of the exosystem and thereby achieves perfect output regulation. Our approach leverages methods from internal model based controller synthesis, multi-agent consensus over directed networks, and stability of time-varying linear systems; the derived result is an adaptation of the (centralized) internal model principle to the distributed, networked setting.

Satoshi Kawamura, Kai Cai, Masako Kishida

We study a multi-agent output regulation problem, where not all agents have access to the exosystem's dynamics. We propose a fully distributed controller that solves the problem for linear, heterogeneous, and uncertain agent dynamics as well as time-varying directed networks. The distributed controller consists of two parts: (1) an exosystem generator that locally estimates the exosystem dynamics, and (2) a dynamic compensator that, by locally approaching the internal model of the exosystem, achieves perfect output regulation. Moreover, we extend this distributed controller to solve an output synchronization problem where not all agents initially have the same internal model dynamics. Our approach leverages methods from internal model based controller synthesis and multi-agent consensus over time-varying directed networks; the derived result is a generalization of the (centralized) internal model principle to the distributed, networked setting.

Yuhang Wang, Yuling Shi, Mo Yang et al.

LLM agents have demonstrated remarkable capabilities in software development, but their performance is hampered by long interaction contexts, which incur high API costs and latency. While various context compression approaches such as LongLLMLingua have emerged to tackle this challenge, they typically rely on fixed metrics such as PPL, ignoring the task-specific nature of code understanding. As a result, they frequently disrupt syntactic and logical structure and fail to retain critical implementation details. In this paper, we propose SWE-Pruner, a self-adaptive context pruning framework tailored for coding agents. Drawing inspiration from how human programmers "selectively skim" source code during development and debugging, SWE-Pruner performs task-aware adaptive pruning for long contexts. Given the current task, the agent formulates an explicit goal (e.g., "focus on error handling") as a hint to guide the pruning targets. A lightweight neural skimmer (0.6B parameters) is trained to dynamically select relevant lines from the surrounding context given the goal. Evaluations across four benchmarks and multiple models validate SWE-Pruner's effectiveness in various scenarios, achieving 23-54% token reduction on agent tasks like SWE-Bench Verified while even improving success rates, and up to 14.84x compression on single-turn tasks like LongCodeQA with minimal performance impact.

Kaiwen Cai, Xinze Liu, Xia Zhou et al.

The generation of realistic LiDAR point clouds plays a crucial role in the development and evaluation of autonomous driving systems. Although recent methods for 3D LiDAR point cloud generation have shown significant improvements, they still face notable limitations, including the lack of sequential generation capabilities and the inability to produce accurately positioned foreground objects and realistic backgrounds. These shortcomings hinder their practical applicability. In this paper, we introduce DriveLiDAR4D, a novel LiDAR generation pipeline consisting of multimodal conditions and a novel sequential noise prediction model LiDAR4DNet, capable of producing temporally consistent LiDAR scenes with highly controllable foreground objects and realistic backgrounds. To the best of our knowledge, this is the first work to address the sequential generation of LiDAR scenes with full scene manipulation capability in an end-to-end manner. We evaluated DriveLiDAR4D on the nuScenes and KITTI datasets, where we achieved an FRD score of 743.13 and an FVD score of 16.96 on the nuScenes dataset, surpassing the current state-of-the-art (SOTA) method, UniScene, with an performance boost of 37.2% in FRD and 24.1% in FVD, respectively.

Renyuan Zhang, Kai Cai

Recently we developed partial-observation supervisor localization, a top-down approach to distributed control of discrete-event systems (DES) under partial observation. Its essence is the decomposition of the partial-observation monolithic supervisor into partial-observation local controllers for individual controllable events. In this paper we extend the partial-observation supervisor localization to large-scale DES, for which the monolithic supervisor may be incomputable. Specifically, we first employ an efficient heterarchical supervisor synthesis procedure to compute a heterarchical array of partial-observation decentralized supervisors and partial-observation coordinators. Then we localize each of these supervisors/coordinators into partial-observation local controllers. This procedure suggests a systematic approach to the distributed control of large-scale DES under partial observation. The results are illustrated by a system of automatic guided vehicles (AGV) serving a manufacturing workcell.

Takatoshi Motoyama, Kai Cai

We propose a top-down approach for formation control of heterogeneous multi-agent systems, based on the method of eigenstructure assignment. Given the problem of achieving scalable formations on the plane, our approach globally computes a state feedback control that assigns desired closed-loop eigenvalues/eigenvectors. We characterize the relation between the eigenvalues/eigenvectors and the resulting inter-agent communication topology, and design special (sparse) topologies such that the synthesized control may be implemented locally by the individual agents. Moreover, we present a hierarchical synthesis procedure that significantly improves computational efficiency. Finally, we extend the proposed approach to achieve rigid formation and circular motion, and illustrate these results by simulation examples.