Most autonomous wheelchair research targets indoor, map-based navigation or depends on costly, complex sensor arrays. In a paper posted to arXiv, researchers introduce ChairNav, a framework that adapts a general visual navigation model for outdoor sidewalk travel using only an RGB-D camera and an NVIDIA Jetson compute module on a commercial electric wheelchair.

The system is trained in three stages. It first pre-trains on large-scale cross-embodiment navigation datasets transformed into the wheelchair domain, then mid-trains on a small set of wheelchair demonstrations to close the remaining domain gap. Finally, a post-training stage refines the model from user interventions, personalizing behavior to the individual rider.

The authors report that pre-training reduced lateral average displacement error by 18%, while post-training increased the mean distance between user interventions from 29.2 meters to 57.4 meters. On kilometer-scale urban sidewalk routes, the system operated autonomously for 93.4% of total driving time, suggesting that lightweight, personalized wheelchair autonomy is feasible for everyday use.