Task Allocation Algorithms for Robotics Projects

🚀 New preprint: How many robots should work on which task? Imagine you have a swarm of robots and several tasks to solve in parallel. A natural intuition is: put more robots on the hardest task. But what if that intuition is sometimes wrong? This was great teamwork with lead author Simay Atasoy, based on a great idea of Giovanni Reina, and with a lot of help with the theory part from Tobias Töpfer and Sven Kosub. In our new preprint, “Optimal Scalability-Aware Allocation of Swarm Robots”, we look at a deceptively simple question: Given a fixed number of robots and multiple tasks, how should robots be allocated to maximize overall performance when tasks scale differently with team size? Not all tasks benefit equally from adding more robots: - Some scale linearly (more robots ≈ more progress), - some saturate (extra robots help less and less), - and some even become retrograde (too many robots hurt due to interference). A concrete example: Think of several teams of robots jointly deciding whether an area is mostly black or white (standard collective perception task). For small swarms, it can be optimal to allocate more robots to the easy tasks, because they yield higher marginal gains. Only once the swarm is large enough does it become optimal to shift robots toward the harder tasks. This size-dependent reversal is unintuitive — but predictable. 💡 What we contribute - A polynomial-time algorithm that computes the globally optimal allocation based on marginal performance gains. - A unifying framework covering linear, saturating, and retrograde scalability. - Validation using simple robot swarm simulations, including cases where physical interference causes performance to decline when teams get too large. ⚠️ Important The current approach is centralized and offline: - A central planner knows the scalability curves, - allocation is computed before deployment, - and robots do not reallocate during execution. ➡️ Future work A natural next step is moving toward decentralized, online task allocation, where robots estimate marginal gains locally and adapt allocations on the fly. If you're interested in scalability, collective behavior, or why “more robots” is not always better, check the preprint: https://lnkd.in/dAEv_X5d #SwarmRobotics #MultiRobotSystems #CollectiveBehavior #RobotSwarms #Scalability #TaskAllocation #DistributedSystems #CollectiveIntelligence #RoboticsResearch #ai #ml

🤖🤖 How to leverage large language models to enable efficient and reliable long-horizon task allocation and planning for heterogeneous robot teams from natural language commands? We're thrilled to share our recent paper, “LaMMA-P: Generalizable Multi-Agent Long-Horizon Task Allocation and Planning with LM-Driven PDDL Planner,” published in the 2025 IEEE International Conference on Robotics and Automation (ICRA). Our code has been released! We develop a novel framework that integrates the powerful reasoning capabilities of large language models with traditional PDDL-based heuristic planners. LaMMA-P tackles the challenges of long-horizon planning by accurately decomposing complex tasks into manageable subtasks, efficiently allocating them among diverse robots, and synthesizing executable plans across various household scenarios. To evaluate the effectiveness and robustness of our method, we introduce MAT-THOR, a benchmark of multi-agent complex long-horizon tasks based on the AI2-THOR simulator. Our method achieves SOTA performance on MAT-THOR, demonstrating a 105% higher success rate and 36% greater efficiency compared to existing LM-based multi-agent planners. Huge thanks to all my students and collaborators: Xiaopan Zhang, Hao Qin, Fuquan Wang, and Yue Dong! LaMMA-P: Generalizable Multi-Agent Long-Horizon Task Allocation and Planning with LM-Driven PDDL Planner 🔗 Project website: https://lamma-p.github.io
 📜 Paper: https://lnkd.in/gX2XXjD2 ✅ Code: https://lnkd.in/gXgKgKhe #ICRA2025 #artificialintelligence #robotics #robots #taskplanning #multirobot #languagemodels #LLM #longhorizon #PDDL #taskplanning #machinelearning #ai #deeplearning #ml

Reinforcement Learning for Multi-Robot Task Allocation! Coordinating heterogeneous robots to complete tasks efficiently is a major challenge in robotics. Centralized scheduling methods are slow, and traditional reinforcement learning struggles with cooperation and deadlocks. This research introduces a reinforcement learning-based framework for multi-robot task allocation and scheduling, enabling decentralized agents to dynamically form teams and minimize idle time. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀: 1. Attention-Based Coordination: Robots learn task dependencies and adapt their schedules in real time. 2. Constrained Flash forward Mechanism: Prevents deadlocks by guiding agent decisions and improving cooperative planning. 3. Decentralized Multi-Agent RL: Scales to large problems, avoiding the bottlenecks of mixed-integer programming (MIP) solvers. 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁? The framework achieves near-optimal task allocation, outperforming heuristic and optimization-based methods while being 100x faster.  It successfully scales to 150 robots and 500 tasks, demonstrating real-world potential for applications like search and rescue, logistics, and industrial automation. Kudos to Weiheng DAI, Utkarsh Rai, Jimmy Chiun, Yuhong Cao, and Guillaume Sartoretti! 🔗 Read the full paper: https://lnkd.in/gSER3_-D I post the latest and interesting developments in robotics - 𝗳𝗼𝗹𝗹𝗼𝘄 𝗺𝗲 𝘁𝗼 𝘀𝘁𝗮𝘆 𝘂𝗽𝗱𝗮𝘁𝗲𝗱! #ReinforcementLearning #MultiRobot #TaskAllocation #AI #Robotics #Automation #DeepLearning #Optimization