Andreu Matoses Gimenez

Portrait of Andreu Matoses Gimenez

I am a PhD candidate in planning for robotics at the Department of Cognitive Robotics (CoR), TU Delft (Netherlands). My research currently focuses on using generative models and physics simulation for task and motion planning problems. I am supervised by Professor Javier Alonso-Mora and Professor Christian Pek.

Before my PhD, I obtained my bachelor's degree in aerospace engineering from Universitat Politècnica de València (Spain), and a master's degree from KTH Royal Institute of Technology (Sweden). During my MSc, I worked as a research engineer at KTH.

A.MatosesGimenez@tudelft.nl
  • November 2026 Starting in November, I will be a visiting researcher at Princeton University in the group of Prof. Tom Silver.

Research

Flow Policies as Actions of Skill-Level World Models: Learned and Symbolic Abstractions for Long-Horizon Planning

Andreu Matoses Gimenez, Andrei-Carlo Papuc, Chris Pek, Javier Alonso-Mora ·

Submitted to the IEEE International Conference on Robotics and Automation (ICRA) 2027

A world model that steps over whole skills: each action is the input of a flow-matching policy, so one policy execution is one world-model transition. We compare four action abstractions, from a compressed noise seed to a symbolic label, under one world model and one planner, on block rearrangement tasks of up to 14 skills.

project page arxiv poster

Learning Task and Motion Plans from Real Demonstrations with Hybrid Flow Matching

Zuleika Redondo Garcia, Andreu Matoses Gimenez, Javier Alonso-Mora ·

Submitted to the IEEE International Conference on Robotics and Automation (ICRA) 2027

One network generates the symbolic plan and the motion of a mobile manipulator with hybrid flow matching, learned from 109 teleoperated demonstrations. Plan suffixes let the robot replan after every action, and object permutations make the planner goal-conditioned without new demonstrations.

project page arxiv models

Cross-Entropy Optimization of Physically Grounded Task and Motion Plans

Andreu Matoses Gimenez, Nils Wilde, Chris Pek, Javier Alonso-Mora ·

IEEE Robotics and Automation Letters (RA-L) 2026 · Presented at IROS 2026

We use a parallelized physics simulator and cross entropy optimization to find optimal realizations of task and motion planning (TAMP) problems. This allows us to consider dynamics, contacts and the low level controllers that the real robot has, and their impact on the optimality and feasibility of the solutions.

project page ieee xplore paper arxiv poster code (mujoco warp)

SADCHER: Scheduling using Attention-based Dynamic Coalitions of Heterogeneous Robots in Real-Time

Jakob Bichler, Andreu Matoses Gimenez, Javier Alonso-Mora ·

IEEE Int. Symposium on Multi-Robot and Multi-Agent Systems (MRS) 2025

Sadcher is a real-time, imitation-learned task assignment framework for heterogeneous multi-robot teams with dynamic coalitions and task precedence. It predicts robot-task rewards using graph attention and transformers, then applies relaxed bipartite matching to produce feasible, high-quality schedules that scale and outperform learning/heuristic baselines; we also release a dataset of 250k optimal schedules.

website arxiv code (pytorch) dataset msc thesis

Decentralized Aerial Manipulation of a Cable-Suspended Load Using Multi-Agent Reinforcement Learning

Jack Zeng, Andreu Matoses Gimenez, Eugene Vinitsky, Javier Alonso-Mora, Sihao Sun ·

Conference on Robot Learning (CoRL) 2025. Best poster award at MACRAI workshop, IROS 2025.

Decentralized MARL for 6-DoF control of a cable-suspended load with multiple MAVs, requiring no inter-drone communication and enabling scalable, onboard deployment. Combined with low-level controllers, it transfers robustly from simulation to reality, matching centralized performance while tolerating uncertainties and even loss of a drone.

website corl paper arxiv code video

Linear Temporal Logic (LTL) Planner for Agricultural Robotics

Shankar Deka, Sujet Phodapol, Andreu Matoses Gimenez, Victor Nan Fernandez-Ayala, Rufus Cheuk Yin Wong, Pian Yu, Xiao Tan, Dimos V. Dimarogonas ·

IEEE International Conference on Automation Science and Engineering (CASE) 2024

Implemented a ROS based linear temporal logic (LTL) planner for the CANOPIES ERC: Collaborative Paradigm for Human Workers and Multi-Robot Teams in Precision Agriculture Systems.

project page paper erc website code

Scalable Multi-Agent Reinforcement Learning for Collision Avoidance

Andreu Matoses Gimenez ·

Master's Thesis, KTH Royal Institute of Technology

Scalable multi-agent reinforcement learning for formation control with collision avoidance. This work was my Master's Thesis while at KTH Royal Institute of Technology. The proposed method exploits the reward structure to enable local approximation of Q-functions and policy gradients, allowing for scalable training. We compare discrete and continuous policies and analyze the impact of the sensing radius on performance and collision avoidance.

project page paper code

ALPHA, a high altitude UAV

Victor Nan Fernandez-Ayala, László Vimláti, Helena Delmotte, Andreu Matoses Gimenez, Mykola Ivchenko, Raffaello Mariani ·

International Conference on Aeronautical Sciences (ICAS) 2022

Design of an autonomous high altitude long endurance UAV to study optical phenomena in the upper atmosphere using scientific imaging instruments. Project under the KTH Space Physics department. Paper presented at ICAS 2022.

project page paper official website