Embedded Files
I2FORD Lab
Theses related to battery disassembly and, in general, to the industrial robotic cell I2FORD. Usually more practical projects.
Automatic Control for Intelligent Robotics (CARI) Lab
Theses related to human-robot collaboration: human motion prediction and replanning strategies. Usually more theoretical projects.
SmolVLA: Efficient Vision-Language-Action Model
Background: Video-Language-Action (VLA) models enable robots to map visual observations and language instructions into actions, making them promising for imitation learning with fewer handcrafted components and better task generalization.
Task:
Train and evaluate a VLA-based imitation learning pipeline on manipulation tasks with a UR10e.
Perform a systematic study of performance as the number of demonstrations varies (data efficiency / scaling trends).
Software: ROS2, UR10e driver/SDK, VLA/SmolVLA training & inference stack (PyTorch), MPPI/MPC module (C++/Python), data collection and evaluation tools.
Reference: https://huggingface.co/blog/smolvla
AVAILABLE
Reinforcement Learning using Soft-Actor-Critic
Background: Reinforcement Learning (RL) enables robots to learn control policies through trial-and-error, and Soft Actor-Critic (SAC) is a state-of-the-art off-policy algorithm that is sample-efficient and stable thanks to entropy-regularized learning.
Task:
Design and implement an RL pipeline based on SAC to learn a control policy for a selected robotics task (e.g., mobile navigation or manipulation).
Define state, action, and reward formulation, with safety and task-completion objectives (e.g., tracking error, energy use, collisions/constraints).
Train in simulation with domain randomization and evaluate generalization; optionally transfer to real hardware with conservative fine-tuning and safety constraints.
Benchmark SAC against classical baselines (e.g., PID/MPC) and/or alternative RL methods, reporting learning curves, robustness, and success rates.
Software: ROS2, MuJoCo (simulation), PyTorch.
References:
Haarnoja et al., Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor (and subsequent SAC improvements).
AVAILABLE
Human-aware planning for mobile platform
Background: Human-aware navigation is essential for mobile robots operating in shared spaces. Standard planners often rely on static costmaps, which limits their ability to anticipate human motion and social constraints.
Task:
Develop an MPPI-based motion planning/control framework for a mobile robot that incorporates social costmaps (e.g., personal space, interaction norms) alongside geometric obstacles.
Integrate a module to predict the temporal evolution of the social costmap, enabling proactive, risk-aware trajectory optimization rather than reacting to instantaneous snapshots.
Ensure the approach runs online and remains compatible with typical ROS2 navigation/control pipelines.
Software: ROS2, MPPI controller implementation, social costmap generation, human motion prediction module (C++/Python as appropriate).
References: Dynamic Risk-Aware MPPI for Mobile Robots in Crowds via Efficient Monte Carlo Approximations
AVAILABLE
Estimation of dynamic model of mobile manipulator considering uncertainties
Background: Dynamic models of manipulators are critical for control, sensing and simulation. They are normally estimated using a deterministic approach based on Least-Square optimization. However, the knowledge of model uncertainties can strongly improve algorithm performance.
Task:
Usage of regression for parameter estimation of the dynamic model
Usage of neural network (NN) to model complex dynamic.
Integration with the existing library
Experimental campaign on a mobile platform
Software: Python
References:
AVAILABLE
Trajectory planning for surface operations
Background: Pollishing, 3D printing, and similar task require to reach an high number of poses on a surface, commonly allowing operational redundancy. This trajectory should be computed to minimize time or energy consumption.
Task:
Using RRT-based algorithm to compute the path.
Integrate MPC-based algorithm to compute the trajectory.
Experimental campaign on a UR10e
Software: C++ (mainly ROS2)
References:
Bezier lib
AVAILABLE
Integration of EtherCAT drives into the ROS 2 framework
Background: The I2FORD cell includes a 2-axis cart used to move the battery in the cell and to allow the robots to reach every side of battery.
Task:
Develop a high-level interface for the battery handler cart compatible with ROS2. In the ros2_control jargon, this will be the hardware_interface module.
This hardware interface should be compatible with the joint_trajectory_controller
The interface will communicate with a low level drive, in control of two axis (one to control the height of the battery, the other to control the inclination)
Communication will happen through EtherCAT bus.
Software: ROS2, ros2_control, EtherCAT ROS2 Driver, C++
References:
Coming soon
Safety aware MPPI
Background: Speed and Separation Monitoring (SSM) is proposed in ISO/TS-15066 as a safety method for human-robot collaborative cells. It requires a minimum separation distance between the human and the robot to avoid collisions, which translates into a maximum robot velocity based on the current position and velocity of both.
MPPI is a novel "MPCish"-like algorithm based on stochastic optimization.
Task:
Implementation of MPPI algorithm integrated with SSM algorithm for computing trajectory following of a mobile platform
Software: C++ (Python)
References:
DONE
SSM-aware replanning strategy using elastic strips
Background: Speed and Separation Monitoring (SSM) is proposed in ISO/TS-15066 as a safety method for human-robot collaborative cells. It requires a minimum separation distance between the human and the robot to avoid collisions, which translates into a maximum robot velocity based on the current position and velocity of both.
Elastic Strip is an obstacle avoidance method that adds an elastic term to the control action to allow temporary deviations from the original trajectory.Task:
Include SSM computations into the Elastic Strip algorithm. The idea is to exploit the elastic strip to bend the trajectory to maximize the SSM and thus avoid slowdowns.
Software: MATLAB (C++, Python)
References:
DONE
Replanning with pose constraints
Background: The replanning procedure in a robot consists of computing a new trajectory in real time due to changes in the environment (new or moving obstacles, etc.)
Task:
Including pose constraints in the MARS replanner to plan the new path while keeping a predefined orientation of the tool.
Software: MATLAB (C++)
References:
DONE
Estimation of dynamic model of robot considering uncertainties and PINN
Background: Dynamic models of manipulators are critical for control, sensing and simulation. They are normally estimated using a deterministic approach based on Least-Square optimization. However, the knowledge of model uncertainties can strongly improve algorithm performance.
Task:
Usage of regression for parameter estimation of the dynamic model
Usage of physical informed neural network (PINN).
Integration with the existing library
Experimental campaign on a UR10e
Software: Python (mainly) and C++
References:
DONE
Page updated
Report abuse