Assistive Robotics
for Citizens

User modelling in Socially Assistive Robotics: two case studies 

Older adults and children are the preferred users of Socially Assistive Robots, typically in the context of applications aiming to retain/foster their cognitive capabilities. In such cases, user modelling is key to ensure the effectiveness of the robot, yet complicated by the dynamic nature of those two phases of life. In this talk, I will discuss the approaches adopted for user modelling in two projects I have been involved in, one targeting older adults and the other primary school children. I will briefly discuss their characteristics, pros and cons and outcomes, and share lessons, fun facts and unexpected challenges that I have encountered along the way.

Structured robot learning in the human-centred design of intelligent assistants 

Endowing Robots with the Ability of Intention Recognition and Decision Making

Recognizing human intentions is considered a crucial ability of assistive robots to endow them with predictive capabilities. This allows them to not only adapt to the current situation, but also to anticipate human behavior and adjust own decisions accordingly. In this talk, we will emphasize typical challenges faced in developing robot intention recognition and decision making systems, whereby specific focus will be given to systems for the recognition of human actions and plans as well as for decision making behind the orchestration of actions, both inspired from humans.

Assistive sensorimotor augmentation as a new paradigm of human-robot collaboration   

Sensorimotor augmentation re-defines the nature of physical human-robot interaction, as it envisages to integrate humans and AI-powered supernumerary robotic limbs to perform complex manipulation tasks with both biological and artificial limbs under the direct control of the human. A fundamental challenge is to find the right trade-off between motion task parameters that are controlled by the user, and the level of autonomy of the robot. The enabling core technologies of this interplay are wearable sensorimotor interfaces that establish a connection between the human sensorimotor system and the system of actuators and sensors of the robot, allowing for reciprocal awareness, trustworthiness, and mutual understanding. So far, wearable extra limbs have been mainly used to augment healthy humans’ capabilities, while collaborative arms and grippers have been mainly applied to industrial assembly tasks. Assistive sensorimotor augmentation, instead, focuses on applying these technologies as assistive devices and targets a large set of end-users with upper-limb disabilities.

Leveraging structure in Robot-Assisted Feeding    

In this talk, I will discuss the role of incorporating structure in developing robot policies that can advance the field of assistive robotics. I will first discuss our work in learning latent actions – low dimensional action spaces that enable intuitive teleoperation of robots through shared autonomy. I will then discuss another type of dimensionality reduction, i.e., role assignment for achieving bimanual manipulation tasks. Specifically, we learn a keypoint-conditioned stabilizing policy and a separate acting policy through imitation learning that can enable substantial generalization in low data regime settings for bimanual tasks. I will then switch to discussing the application of assistive feeding and how that can leverage from similar structures. This includes bite acquisition strategies such as bimanual scooping, skewering food using both haptic and visual information, and effectively combining different low-level bite acquisition primitives for long-horizon food manipulation such as switching between twirling and skewering spaghetti to clear a plate. I will end the talk with briefly discussing the challenges of in-mouth bite transfer that can build upon our bite acquisition strategies.