“MIRA”: A Multitasking Intelligent Robotic Assistant Supporting Activities of Daily Living

Le résumé fourni par la source

Assistive robots are increasingly proposed as partners in everyday life, yet existing systems remain fragmented across mobility, communication, and task execution, and rarely learn from the people they are designed to help. Caregivers in hospitals, rehabilitation centers, and assisted living facilities face mounting workloads while supporting individuals whose needs span walking assistance, conversational interaction, and personalized daily routines. Meeting these needs requires a unified assistive system rather than a collection of isolated capabilities, and one that adapts to the individual user through ordinary interaction rather than burdensome calibration. This Thesis presents the design, implementation, and evaluation of MIRA, a Multitasking Intelligent Robotic Assistant for activities of daily living. MIRA is built on a Franka Emika Panda manipulator mounted on a Clearpath Ridgeback omnidirectional base and integrates three core capabilities within a single platform. First, a shared-control reinforcement learning framework enables human–robot co-navigation during walking assistance, jointly satisfying patient support, pedestrian avoidance, and static obstacle navigation in crowded indoor environments. Second, the Speech2Action framework translates natural spoken intent into executable robotic behavior through a two-stage pipeline that decouples utterance classification from command generation. Third, a Markov chain-based preference learning framework reframes everyday substitution events, such as missing ingredients during cooking, as structured preference queries, combining a local user-specific model with a universal population-level prior to support both cold-start and long-term personalization. The proposed frameworks were evaluated on the physical MIRA platform across simulated and laboratory settings. The shared-control navigation framework achieved up to a 90% success rate in environments containing both dynamic and static obstacles while preserving patient contact throughout navigation. The Speech2Action framework reached approximately 88.3% robot execution accuracy on the physical platform; an accompanying user study with 35 participants further revealed that command length, politeness, and sentiment are systematically shaped by perceived qualities of the robot. The Markov-theoretic preference learning framework was evaluated in a three-visit user study with 16 participants: acceptance of the robot’s first-ranked substitute increased from 31.81% to 65.38% between visits, System Usability Scale and suggestion satisfaction scores increased significantly under repeated missing-item conditions (𝑝 = 0.01 and 𝑝 = 0.03, respectively), and held-out cross-validation achieved a 73.93% hit rate, indicating that the learned preference structure generalizes beyond the specific individuals from whom it was collected. Together, these contributions demonstrate that a single robotic assistant can guide a user through a crowded environment, understand a casually phrased request, and learn individual preferences from a small number of natural interactions. The Thesis advances the broader proposition that adaptation is a unifying design principle for assistive robotics rather than an isolated feature, that real-world constraints can be reframed as informative signals rather than failures, and that meaningful personalization is achievable within the data environment of a single household or small assistive care facility.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.