Context

Vine pruning is an essential but time- and labor-intensive viticultural operation. Automating this task with a robotic system requires fine-grained perception of the vine stock (structure, cane positions, pruning points) as well as a decision-making policy capable of adapting to the variability of individual plants. Reinforcement learning (RL) is a promising approach for learning robust pruning strategies from sensor data, 2D vision, and/or 3D point clouds.

• Duration of the internship: 4 to 6 months (final-year/graduation internship)
• Required level: Master's degree — 2nd year Master / final-year engineering school
• Location: LAAS-CNRS Toulouse, France
• Starting date is flexible depending on the candidates (first or second semester)

Internship Objectives

This internship is part of a research project exploring and comparing several RL approaches applied to a robotic manipulator for vine pruning:

• Vision-based RL: learning from 2D images (RGB / RGB-D) of the vine stock
• Point cloud-based RL: learning from 3D representations of the plant
• Multimodal fusion: combining the vision and point cloud approaches (and potentially sensors) into a unified policy

Given the scope, the internship may focus on a priority subset of these components (e.g. vision + point cloud + fusion), with a gradual extension to the full scope depending on progress.

Main Tasks

• Literature review on RL methods applied to agricultural/robotic perception and manipulation
• Setting up a learning environment (simulation and/or real data) for the vine pruning task
• Implementing and training RL policies for each modality (sensors, vision, point cloud)
• Designing a fusion strategy combining the vision and point cloud modalities
• Comparative evaluation of the approaches (pruning accuracy metrics, robustness, convergence time, etc.)
• Writing an internship report and presenting results to the team

Candidate Profile

Master's student (2nd year) or final-year engineering student, specializing in robotics, computer vision, AI, or machine learning with:

• Solid foundations in reinforcement learning
• Experience with Python and deep learning frameworks (PyTorch preferred)
• Autonomy, scientific rigor, and enthusiasm for experimentation
• Knowledge of point cloud processing (Open3D, PCL, PointNet, etc.) and/or computer vision is a plus
• Prior experience in robotics (simulation, ROS) is a plus

Supervision

The intern will be supervised by Fadi Gebrayel, Martin Mujica and Patrick Danes and will join the RAP team in the LAAS-CNRS.

How to Apply

Please send your CV, cover letter and academic transcripts of the last 2 years to