Personalizable Partial Automation

Watch the CHI PLAY presentation →

Some players cannot control games using standard hardware, regardless of how that hardware is adapted. The inputs required may simply be unavailable to them. Partial automation addresses this by delegating inaccessible inputs to an AI partner while the player retains control over everything they can manage. The AI fills in for the player's specific limitations; the player remains an active participant.

I built partial automation systems for two games — a cycling rehabilitation game and a shooting game — and tested them with six participants with spinal cord injuries of significantly different severities. Participants were recruited through a sustained volunteer relationship with Spinal Cord Injury Ontario, developed alongside formal accessibility training through the NSERC CREATE READi program. All six participants could play both games. This result held across a range of motor profiles that included participants who could produce almost no voluntary movement.

The personalization dimension was important. Participants strongly preferred familiar control devices — joysticks configured like their wheelchairs, sip-and-puff controllers they already used — over devices optimised for able-bodied players. The AI's adaptation to each player's specific motor profile, rather than applying a standard intervention, made the difference between systems participants would use and systems they would not.

The study also produced the first clear evidence of the task-demand inference finding. Several participants attributed to themselves actions the AI was performing — steering the avatar, firing the weapon — because those were the actions they wanted to be doing. The inference ran from desire and task goal, not from observation of the interface. These participants were not confused about the controls. They were interpreting the system through the lens of what they intended to accomplish within it.

Read the paper →