Abstract
PEACE [1] is a planner–executor agent for a single-robot systems: a large language model (LLM) turns a natural-language mission into one typed plan of tool calls, an executor dispatches it deterministically over ROS 2, and an independent constraint layer validates every motion command. Many missions that motivate language-tasked drones, such as area survey, search, or inspection of distributed infrastructure, are naturally multi-robot. This project extends the planner–executor paradigm from single-robot system to a multi-robot system: the planner must decompose a mission into per-robot plans with dependencies, the executor must coordinate several robots, the world model must be shared, and the constraint layer must enforce inter-robot constraints such as separation and operational environment partitioning in addition to per-robot geofences.
Supervision
- Erdem Uysal
Motivation
Language interfaces for swarms have so far taken one of two routes. Either the LLM emits a global shape or objective that a reactive swarm controller follows, as in FlockGPT [2] and language-guided pattern formation [3], or each robot runs its own LLM and the robots negotiate in natural language, as in LLM2Swarm [4]. The first route cannot express heterogeneous, multi-step missions; the second keeps an LLM in every robot’s control loop, multiplying latency, cost, and hallucination risk. Heterogeneous-team planners such as SPINE-HT [5] show that an LLM can generate a dependency graph of grounded subtasks and assign them to robots by capability, but they run a planning iteration at every step, are centralized, and treat the aerial robot as a waypoint server rather than as a planning agent in its own right.
The planner–executor paradigm suggests a third route: one planning pass produces a set of typed per-robot plans plus the dependencies between them, which a coordinating executor dispatches without further model calls, replanning only when a robot reports a tool failure. All of PEACE’s benefits then carry over to the swarm: the action space stays fixed and inspectable, plans remain auditable per robot, and safety is enforced outside the model. The open questions are how to represent multi-robot plans so that they stay typed and verifiable, how to allocate subtasks without relying on the LLM’s weak spatial reasoning [5], how to merge per-robot detections into one consistent world model, and how to extend the constraint layer to inter-robot separation and to airspace partitioning between robots.
Goal
The student will design, implement, and evaluate a swarm extension of PEACE in PX4 multi-robot software-in-the-loop simulation with Gazebo:
- Multi-robot plan representation. Extend PEACE’s typed MissionPlan to a set of per-robot plans with explicit dependencies (a directed acyclic graph of tool calls), so that independent subtasks execute in parallel and dependent ones wait for reported results.
- Allocation. Separate what the LLM decides (which subtasks are needed, their order, their semantic targets) from what a deterministic allocator decides (which robot flies where), using distance, battery, and payload as costs.
- Shared world model. Merge detections from several robots into one 3D object map with provenance and uncertainty, and expose a consistent snapshot to the planner.
- Swarm-level constraint enforcement. Extend the safety service with minimum inter-robot separation, per-robot geofences and altitude bands, and airspace partitioning, and define what happens when one robot’s abort invalidates another robot’s plan.
- Evaluation on a set of multi-robot missions (survey, multi-object search, distributed inspection) with 2 to 6 robots: mission success, LLM calls, time to completion, constraint violations, and a comparison against a per-robot ReAct-style baseline and against a single-UAV PEACE executing the same mission sequentially.
Requirements
- Strong Python skills; working knowledge of ROS 2; experience with, or willingness to learn, PX4 multi-robot SITL and Gazebo.
- Background in multi-robot systems or task allocation (assignment problems, scheduling) is an advantage.
- Familiarity with LLM APIs and structured output (JSON schemas, tool calling).
- Helpful: experience with distributed systems or message-passing middleware.
Pointers
[1] E. Uysal, T. Kehrer, and S. Panichella, “PEACE: A planner–executor agent with constraint enforcement for UAVs,” arXiv preprint arXiv:2606.00104, 2026.
[2] A. Lykov, S. Karaf, M. Martynov, V. Serpiva, A. Fedoseev, M. Konenkov, and D. Tsetserukou, “FlockGPT: Guiding UAV flocking with linguistic orchestration,” in Proc. IEEE Int. Symp. Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), 2024, pp. 485–488.
[3] H.-S. Liu, S. Kuroki, T. Kozuno, W.-F. Sun, and C.-Y. Lee, “Language-guided pattern formation for swarm robotics with multi-agent reinforcement learning,” in Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS), 2024, pp. 8998–9005.
[4] V. Strobel, M. Dorigo, and M. Fritz, “LLM2Swarm: Robot swarms that responsively reason, plan, and collaborate through LLMs,” arXiv preprint arXiv:2410.11387, 2024.
[5] Z. Ravichandran, F. Cladera, A. Prabhu, J. Hughes, V. Murali, C. J. Taylor, G. J. Pappas, and V. Kumar, “Heterogeneous robot collaboration in unstructured environments with grounded generative intelligence,” arXiv preprint arXiv:2510.26915, 2025.
[6] B. Yu, Q. Yuan, K. Li, H. Kasaei, and M. Cao, “Co-NavGPT: Multi-robot cooperative visual semantic navigation using vision language models,” IEEE Robotics and Automation Letters, vol. 11, no. 2, pp. 2122–2129, 2025.
[7] O. Sautenkov, Y. Yaqoot, M. A. Mustafa, F. Batool, J. Sam, A. Lykov, C.-Y. Wen, and D. Tsetserukou, “UAV-CodeAgents: Scalable UAV mission planning via multi-agent ReAct and vision-language reasoning,” arXiv preprint arXiv:2505.07236, 2025.
PX4 multi-vehicle simulation documentation:
Contact
Erdem Uysal, Institute of Computer Science, University of Bern ()