Presentation
Agency-Constrained Shared Autonomy for Prosthetic Grasp Selection: An Interactive-AI Controller and a Reproducible Offline Benchmark
DescriptionUpper-limb prostheses increasingly integrate autonomy (e.g., context-aware grasp inference from gaze or vision) with user control. Yet, assistance that improves task success can also erode agency—the user’s experience of “being the one who is causing” the action. This workshop paper contributes (i) an agency-constrained shared autonomy (ACSA) formulation that explicitly bounds autonomy-induced disagreement via an Agency Loss Index (ALI), and (ii) a fully reproducible offline interaction benchmark that replays multimodal prosthesis episodes from NinaPro DB10 (5 able-bodied, 4 transradial amputees; 3,814 grasp episodes) to compare arbitration strategies
without new human-subject approvals. Beyond a binary gate (use autonomy only if ALI ≤ 𝜏), we introduce a set-based envelope controller (SetACSA) that selects the autonomy-preferred action within an agency-feasible set. Across subjects, SetACSA improves grasp classification accuracy from 0.532 ± 0.021 (user-only) to 0.580 ± 0.014 while reducing agency loss (0.012 ± 0.001 vs. 0.087 ± 0.004 for confidence-weighted blending). The framework reframes shared autonomy for assistive robotics as an interactive-AI design problem with tunable agency budgets and provides an open artifact for community replication and extension.
without new human-subject approvals. Beyond a binary gate (use autonomy only if ALI ≤ 𝜏), we introduce a set-based envelope controller (SetACSA) that selects the autonomy-preferred action within an agency-feasible set. Across subjects, SetACSA improves grasp classification accuracy from 0.532 ± 0.021 (user-only) to 0.580 ± 0.014 while reducing agency loss (0.012 ± 0.001 vs. 0.087 ± 0.004 for confidence-weighted blending). The framework reframes shared autonomy for assistive robotics as an interactive-AI design problem with tunable agency budgets and provides an open artifact for community replication and extension.
Event Type
Lecture
TimeThursday, October 22nd3:20pm - 3:40pm PDT
Location
Similar Presentations