Context

As part of a collaboration between Siléane and LAAS-CNRS, we are offering a 6-month research internship focused on advancing Physical AI for industrial robotics.

The internship will take place within the MAGMA research project, which investigates how intelligent agents can understand, decompose, plan, and execute complex robotic manipulation tasks over extended time horizons. The project has already led to research contributions, including MAGMA-GEN [1], accepted at the top-tier Conference on Robot Learning (CoRL 2026), and ongoing research on MAGMA-BENCH [2], a framework for evaluating robotic agents on complex and dynamic tasks.

While current approaches show promising capabilities, significant challenges remain in coordinating multiple robots, optimizing task execution, and adapting plans to unexpected events.

This internship aims to investigate novel agent architectures capable of addressing these challenges.

Objectives

The intern will work at the intersection of Agentic AI, Task Planning, and Multi-Robot Systems. The main research question is:

How can intelligent agents combine structured task representations, reasoning capabilities, and classical planning methods to efficiently coordinate multiple robots in dynamicenvironments?

The internship will focus on three main research directions:

1. Agent Architecture & Task Representation
   1. Investigate and propose novel architectures for long-horizon robotic task planning.
   2. Explore structured task representations, including task graphs, hierarchical planning, and symbolic representations.
   3. Study how Large Language Models (LLMs) can interact with classical planning approaches, such as PDDL-based planners, constraint optimization methods, task allocation and scheduling strategies for multi-robot coordination [3].
   4. Investigate how agents can react to unexpected events, execution failures, and changing task requirements.
   5. Evaluate the trade-offs between planning quality, computational cost, and adaptability.
2. Experimental Evaluation & Benchmarking
   1. Use MAGMA-GEN and MAGMA-BENCH to generate scenarios and evaluate proposed architectures.
   2. Design new multi-robot scenarios involving dependencies, shared resources, and dynamic events.
   3. Compare the proposed approach against existing agent architectures and planning baselines.
   4. Validate selected capabilities on real industrial robotic platforms.

The intern will be encouraged to take initiative, explore original ideas, and contribute actively to research decisions, from the initial literature review to the experimental evaluation. The internship may lead to a scientific publication and potential future employment opportunities.

Expected skills

We are looking for a motivated Master's student (M2 or equivalent) in Artificial Intelligence, Robotics, Computer Science, or a related field. We particularly value candidates who have already explored AI through research internships, academic projects, open-source contributions, or personal projects.

Required

• Good proficiency in Python
• Good understanding of Machine Learning and modern AI methods.
• Solid software engineering practices (modular design, clean code, experimentation).
• Ability to understand and implement methods from scientific papers.
• Autonomy, curiosity and creativity.

Highly appreciated

• Experience with LLMs, agentic AI, or LLM fine-tuning.
• Familiarity with task and motion planning, PDDL, or optimization methods.
• Knowledge of reinforcement learning or imitation learning.
• Experience with robotic simulation environments (Isaac Sim, ManiSkill, etc.),
• Familiarity with Git, Linux, and C++.
• Previous experience with robotic systems or multi-agent environments.

Candidates do not need to be expert in all these topics! We are primarily looking for someone eager to explore new approaches, experiment, and tackle challenging research problems.

Practical information

• Level: Master 2
• Duration: 4-6 months.
• Location: Siléane, 42000 Saint-Etienne
• Funding: 1000€/month
• Langage: English or French

Bibliography

1. Loan Bernat, Matthieu Grard, Ariane Herbulot, Florent Lamiraux. MAGMA-GEN: Validated Recovery Supervision from Ambiguous Failures via Counterfactual Re-Execution. 10th Conference on Robot Learning (CoRL 2026), Conference on Robot Learning (CoRL), Nov 2026, Austin (Texas, USA), United States.
2. Loan Bernat, Matthieu Grard, Ariane Herbulot, Florent Lamiraux. MAGMA-Bench: Evaluating Long-Horizon Embodied Agents under Dynamic Tasks and Execution Failures. 2026.
3. Xiangyu Su, Juzhan Xu, Oliver van Kaick, Kai Xu, Ruizhen Hu. IMR-LLM: Industrial Multi-Robot Task Planning and Program Generation using Large Language Models. 2026 IEEE International Conference on Robotics and Automation (ICRA), pp. 7609–7616, June 2026, Vienna, Austria.