Assist-as-needed robots are designed to provide only as much help as a user requires, but that requires knowing why a movement changed. The same observed motion or interaction force could mean the user has a new goal, or that their muscles can no longer produce the same effort. A new paper on arXiv proposes an attribution-gated framework that jointly estimates both intention and capability, rather than treating them as separate problems.
The authors describe a dual-measurand approach, meaning the system tracks two distinct quantities at once. Attribution gating is used to decide which measurement should influence the robot's assistance at any moment, based on whether the change is likely due to a shift in intent or a drop in effective voluntary torque. This is intended to avoid misinterpreting a capability loss as a change of mind, or vice versa.
The work addresses a core challenge in assist-as-needed control for upper-limb rehabilitation and daily assistance. By making the distinction explicit, the method could lead to robots that respond more safely and appropriately to users whose strength and goals fluctuate. The paper is available on arXiv under identifier 2610.21534.