decentralized vision-based control

This work proposes a decentralized aggregation method for multi-agent systems equipped only with limited range and bearing sensing capabilities, leveraging image-based observations as input. Local perceptual information is encoded into structured images, from which spatial features are extracted using convolutional neural networks, and a scalable control policy is trained via deep reinforcement learning. By representing agent perception as images—an approach novel in multi-agent settings—the method circumvents the limitations of handcrafted feature engineering or fixed vector representations, significantly enhancing policy generalization across varying group sizes and initial configurations. Experimental results demonstrate that the proposed approach achieves high success rates in diverse and complex scenarios, exhibits convergence speeds comparable to VariAntNet, and, in several challenging setups, emerges as the sole viable solution, substantially outperforming conventional analytical methods.

This work addresses decentralized multi-agent navigation in cluttered environments, proposing the first joint optimization framework for agent policies and reconfigurable environmental layouts (e.g., obstacle placements). Methodologically, it employs model-free policy gradient reinforcement learning and introduces a two-stage alternating optimization algorithm that concurrently updates distributed agent policies and environmental structure. Theoretical analysis establishes convergence to local minima of a time-varying non-convex optimization problem. A key finding is that the optimized environment autonomously forms implicit, motion-decoupled guidance structures—enhancing behavioral coordination without explicit communication or centralized control. Experiments across diverse dense scenarios demonstrate consistent superiority over baselines in navigation success rate, throughput efficiency, and collision rate, empirically validating that environmental configuration optimization delivers substantial gains for multi-agent collaborative navigation.

This work addresses the challenge of decentralized collaborative perception and navigation for multi-robot systems operating in dynamic, unknown environments under low-bandwidth ad-hoc networking constraints. We propose a lightweight distributed framework that leverages an interpretable information-flow mechanism and an environment uncertainty interaction field model to enable observation sharing, complementary uncertainty fusion, and conflict-free convergent/divergent motion. Furthermore, we design a fully decentralized, training-free, self-organizing path optimization algorithm that requires no central coordinator. The approach significantly reduces path redundancy (average 38% reduction in simulation and real-world experiments), enhances task robustness (52% improvement in success rate under communication disruptions), and incurs minimal computational and communication overhead—enabling real-time coordination among ten or more robots. The framework has been validated on edge computing platforms, demonstrating plug-and-play deployment and scalability.

Large-scale robotic swarms face challenges in collaborative coverage tasks across expansive environments, including tight perception-communication-action (PAC) coupling, rigid synchronous execution paradigms, and poor generalizability of decentralized control policies. Method: This paper proposes the first fully decentralized, asynchronous Graph Neural Network (GNN)-driven framework. It introduces an asynchronous PAC architecture with an adaptive message-aggregation mechanism, enabling each robot to independently perform perception, local communication, and control decision-making on demand. Furthermore, it incorporates hidden-layer information exchange and distributed navigation policy learning to overcome bottlenecks inherent in sequential execution and centralized evaluation. Contribution/Results: Experiments at the thousand-robot scale demonstrate that our approach reduces communication overhead by 40% and shortens task completion time by 35% compared to state-of-the-art baselines, significantly improving coverage efficiency and system robustness.

To address the challenge of collaborative coverage monitoring by decentralized robot swarms in unknown environments under communication and sensing constraints, this paper proposes an end-to-end trainable perception–action–communication closed-loop architecture. The method innovatively integrates a CNN for local observation processing, a GNN for dynamic topology modeling and adaptive communication-content selection and fusion strategy learning, and a shallow MLP for generating distributed control actions—enabling their joint optimization for the first time. Trained via imitation learning, the framework significantly outperforms classical centralized and distributed baselines in coverage efficiency. It exhibits strong generalization (to unseen environments), scalability (seamless deployment on larger swarms), and robustness (stable performance under positional estimation noise).

This work addresses the challenge of enabling agents to implicitly convey internal state information through their actions in communication-constrained environments, thereby facilitating accurate external observation. The authors propose a method that directly embeds state observability into the reinforcement learning reward function, guiding the policy to actively expose informative state signals while preserving primary task performance. By integrating reinforcement learning with observability-aware optimization, the approach successfully trains control policies with high observability in an aerial tracking task. Experimental results demonstrate that the resulting policies significantly enhance the accuracy of state reconstruction by external observers, with negligible degradation to the main task performance.

This work addresses the lack of a scalable, decentralized multi-agent control framework for area coverage tasks that offers theoretical convergence guarantees. The authors propose a finite-state, decentralized policy-based control approach that decouples the coverage problem into two components: a reference-configuration-guided deep neural network design and a distributed control strategy based on agent policies. Central to their method are an anchor-follower triangular communication topology and the novel concept of Anyway Output Controllability (AOC). This framework yields a computationally efficient, time-invariant control policy capable of dynamically adapting to changes in the target region while theoretically ensuring convergence to the optimal coverage configuration, thereby achieving scalable and collaboratively efficient multi-agent coverage control.

This work addresses the challenge of decentralized cooperative pursuit by multiple autonomous aerial vehicles (AAVs) in cluttered environments using only local, noisy sensing. The authors propose an end-to-end multi-agent reinforcement learning framework that directly maps raw LiDAR data to continuous control commands. The key innovation lies in the predictive spatiotemporal observation (PSTO) representation, which unifies the encoding of obstacles, target intent, and teammate motion, thereby integrating perception and decision-making into a single pipeline. Notably, the approach operates without access to privileged information and enables zero-shot policy transfer across different team sizes. Simulations demonstrate superior capture efficiency and success rates compared to existing methods that rely on privileged information. Furthermore, real-world outdoor experiments with a quadrotor swarm validate the system’s fully autonomous, onboard-only capability for cooperative pursuit.

This work addresses the challenge of open-vocabulary, multi-target semantic navigation for multi-robot systems operating without centralized coordination or shared global maps. The authors propose a fully decentralized multi-agent framework that enables efficient autonomous collaboration through peer-to-peer asynchronous communication of local maps and navigation intents. By integrating an implicit task allocation mechanism, a distance-weighted frontier selection strategy, and open-vocabulary semantic understanding, the system eliminates the need for centralized scheduling, thereby reducing redundant exploration and avoiding single points of failure. Experimental results demonstrate that the approach matches or surpasses centralized baselines on the HM3DSem and GOAT-Bench benchmarks, while real-world deployment with two onboard robots in an office environment confirms its practical feasibility.

This work addresses the challenge of balancing long-term strategic coordination with short-term safety and feasibility of actions in multi-agent systems by proposing a Multi-Agent Actor-Critic Model Predictive Control (MA-AC-MPC) algorithm. MA-AC-MPC uniquely integrates the actor-critic reinforcement learning framework with model predictive control, enabling long-horizon optimization of collaborative policies through reinforcement learning while generating short-horizon control commands that respect dynamic constraints and ensure safety. Evaluated on a heterogeneous robotic platform comprising drones and omnidirectional wheeled robots in a pursuit-evasion task, the method achieves a 100% task success rate—substantially outperforming conventional MLP-based approaches, which attain only 60% success—demonstrating its effectiveness and robustness in complex, dynamic environments.