Advances and Innovations in the Multi-Agent Robotic System (MARS) Challenge

TL;DR

The paper surveys the MARS Challenge, a NeurIPS 2025 benchmark that separates planning and control to study embodied multi-agent coordination in heterogeneous robotic systems. It introduces two tracks—Planning, leveraging vision-language-action models for agent selection and high-level planning; and Control, evaluating end-to-end policy execution on multiple arms in realistic simulations—providing a unified evaluation framework with diverse environments like RoboCasa and ManiSkill3. Key contributions include two top planning methods (self-correction with iterative plan refinement and a modular closed-loop framework) and two control approaches (Combo-MoE with a combinatorial expert pool and CoVLA with decentralized, groundings-based collaboration). The findings highlight iterative planning and structured coordination as central mechanisms enabling scalability, while acknowledging limitations from simulation-only evaluation and the need for real-world transfer, which motivates future work on perception, actuation, and real-time decision-making in embodied MAS. The work advances practical understanding of how to design scalable, robust multi-agent robotic systems using VLMs and modular architectures, with potential impact on future collaborative AI systems and human-agent interaction.

Abstract

Recent advancements in multimodal large language models and vision-languageaction models have significantly driven progress in Embodied AI. As the field transitions toward more complex task scenarios, multi-agent system frameworks are becoming essential for achieving scalable, efficient, and collaborative solutions. This shift is fueled by three primary factors: increasing agent capabilities, enhancing system efficiency through task delegation, and enabling advanced human-agent interactions. To address the challenges posed by multi-agent collaboration, we propose the Multi-Agent Robotic System (MARS) Challenge, held at the NeurIPS 2025 Workshop on SpaVLE. The competition focuses on two critical areas: planning and control, where participants explore multi-agent embodied planning using vision-language models (VLMs) to coordinate tasks and policy execution to perform robotic manipulation in dynamic environments. By evaluating solutions submitted by participants, the challenge provides valuable insights into the design and coordination of embodied multi-agent systems, contributing to the future development of advanced collaborative AI systems.

Figures (7)

• Figure 1: Top 10 Scores
• Figure 2: Top 20 Hardest Tasks
• Figure 3: Visualization of Control Track tasks, including four tasks covering 2, 3, and 4 robotic arms.
• Figure 4: Overall pipeline of the Self-Correction framework. Starting from $N$ manually annotated examples as prompts, VLMs generate seed training data from task instructions and scene observations. Generated plans are evaluated by a judging VLM and refined via supervised fine-tuning using the best plans. This iterative process improves planning performance, while multiple inferences and a voting mechanism address non-unique solutions, resulting in a scalable planning framework.
• Figure 5: The multi-agent planning system comprises three collaborative components: Activate Agent, Planning Agent and Monitor Agent. Upon receiving a user instruction and a scene image, the system selects the appropriate robots and generates a step-by-step execution plan. Specifically, the Activate Agent and Planning Agent are supervised fine-tuned using the L1 (activation tasks) and L2 (planning tasks) datasets, respectively derived from the adjusted VIKI benchmark kang2025viki.