Topics

Core research

User studies are the yardstick to measure the success of our ideas: planning and carrying out tests on users, e.g., amputees, stroke survivors, etc.

Thinking a user study requires knowledge about statistics, functional assessment, your own equipment, and most importantly… humaneness.

Acceptance (user satisfaction and usefulness) is the primary criterion; performance must show correlation with acceptance over time.

Co-adaptation is, informally speaking, reciprocal adaptation of user and machine along time.

Presented with a new tool, a user will learn how to use it and, during the process, learn new motor skills; now, whereas standard “machines” cannot reciprocally adapt to the user, machine learning can and, we claim, should, leading to optimal user control, acceptance and satisfaction.

Measuring and exploiting co-adaptation is still a largely unexplored realm of Human-Machine Interaction.

Intent detection in one sentence: take signals from the user’s body, turn them into actions for an assistive device using machine learning. Typical case: eight electromyography sensors, linear regression, flexion / extension of a prosthetic wrist.

Interactive learning is machine learning which can be updated online and on demand, whenever required by a novel situation, or requested by the user, or… any other reason.

Everything we know, we have constructed it in our own head, out of our experiential landscape, through the interaction with our environment.

This formidable sentence characterises Radical Constructivism, a simple yet powerful paradigm to build artificial beings which could… learn from users and improve acceptance and control of assistive devices.

We record the body in several ways at once: the electrical activity of muscles (surface and high-density EMG), their mechanical behaviour (ultrasound, force myography) and the movement that results (inertial sensors, data gloves, motion tracking). Each modality sees a different part of the picture; together they describe what the user is doing, and intends to do.

Signal processing turns these raw streams into something a learner can use: filtering and feature extraction, dimensionality reduction, synchronisation and fusion across modalities, all in real time. Ultrasound is a current focus, since it images muscle deformation directly, reaches deep muscles and does not suffer from electrode shift.

Is advanced AI really necessary in assistive robotics? Can LSTM, CNNs, Large Language Models and Autoencoders boost the acceptance and performance of a prosthetic arm, for example?

We investigate the usage and design of contemporary AI systems for rehabilitation and prosthetics, with an eye to Human-Machine Interaction.

Application areas

ULP is, traditionally, AIROB’s core research and focusses upon designing and testing the interface between a user, either amputated or born without a limb, and a prosthetic shoulder / wrist / hand / fingers, either virtual or real.

The problem of properly, reliably and dexterously controlling a ULP device is still totally open, notwithstanding 40 years of research, to the point that, for trans-radial amputees, the best solution is still a body-powered hook. Myoelectric prostheses, i.e., employing advanced sensors and machine learning, are still by and large unreliable.

At AIROB, we strive to improve the sensors and the ML methods, but most importantly, the interaction between user and prosthesis (see also co-adaptation).

Lower-limb amputees report that their prostheses are most limiting in everyday situations beyond level walking, such as stairs, slopes, uneven terrain and sit-to-stand transitions, which cause fatigue, fear of falling and reduced trust in the device. Addressing these user-reported challenges is the main motivation of AIROB’s research in this field.

A key requirement is that the prosthesis recognises the user’s intention and anticipates transitions between activities. We investigate intention detection through sensor fusion of surface EMG, kinematic and load data, supported by simulation to develop and safely evaluate new strategies.

The detected intention is translated into joint behaviour through position, velocity and torque control of the knee and ankle, aiming at responsive, reliable and safe operation. This work builds on AIROB’s own powered knee–ankle prosthesis, redesigned in-house on the basis of the Open-Source Leg (OSL) 2.0 to meet the demands of everyday activities (see also intent detection and co-adaptation).

Stroke is a major cause of death and long-term disability worldwide, creating the need for rehabilitation technologies. Frequently, motor function results impaired, making rehabilitation and assistance essential for restoring activities of daily living and independence.

Can rehabilitation technologies, including robotic systems, extended reality (XR) and functional electrical stimulation, support function recovery after stroke? We develop and evaluate methods to make such technologies adaptive to their users.

Our robotic setup, consisting of two Franka Research 3 robotic arms and two Haption input devices with force feedback, allows us to set up experiments in teleoperation involving a large workspace and 6 degrees of freedom. This, coupled with our vast array of muscular activity sensors, enables the study of teleimpedance, i.e., the control of a robotic manipulator on the part of a pilot in terms of desired position, forces, and mechanical stiffness. Multiple force/torque sensors enable the precise measurement of interactions between the pilot and the robots.

Functional Electrical Stimulation is a technique that allows to cause the contraction of muscles through the stimulation of peripheral motor nerves by injecting stimulation currents. Various devices, some of which are internally developed, such as the MyoCeption, can be combined with our kinematiks sensors and machine learning frameworks to assist movement and directed force exertion in the peri-personal space. High-density matrices can also be included in the loop to allow the control of finer muscle groups or to adapt to varying kinematics.

Rehabilitation, especially for prosthesis users, is repetitive and demanding, and its effectiveness is influenced by the patient’s motivation and the amount of training actually completed. Extended reality (XR) environments can help in several ways: they can increase engagement through immersive, interactive settings, enable objective data collection, and support training at home.

Our goal at AIROB is to develop XR environments that support motor learning, simulate real-world activities, and quantify training results, with the aim of connecting virtual practice to everyday life.