BlindPilot: A Robotic Local Navigation System that Leads Blind People to a Landmark Object
Seita Kayukawa, Tatsuya Ishihara, Hironobu Takagi, Shigeo Morishima, Chieko Asakawa · 2020 · Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems (CHI EA '20, Late-Breaking Work) · doi:10.1145/3334480.3382925
Summary
Most indoor-navigation research for blind travellers has focused on what the authors call 'global' navigation — getting from a building entrance to a room, gate, or exhibit. The 'last few metres' of the journey, a body-scale interaction with a specific landmark object such as an empty chair, a door handle, an elevator button, or a counter, is often left to the user's white cane and the systems that deliver voice cues like 'two o'clock, 0.7 metres'. The authors argue that this local-navigation gap imposes real effort: audio-cued approach requires frequent orientation corrections and a zig-zag walking pattern that feels insecure. BlindPilot is a prototype robotic alternative. A ZED stereo camera paired with YOLOv3 detects chairs and nearby people, filters out chairs within 1 m of a person (treating them as occupied), and marks the remaining empty chair as the target. A mobile robot with a Hokuyo LiDAR builds a 2D map via ROS gmapping-based SLAM, generates a path to a point 0.7 m to the left of the chair (so the user can sit naturally), and physically pulls the user along via a handle, at a maximum speed of 1.2 m/s. The study compared BlindPilot to an audio-only baseline modelled on a prior Google Glass + TTS system ('right'/'left'/'straight X m' once per second, delivered from a ZED camera worn on the chest) with six blind cane users approaching an empty chair placed 6 m away at three angles (front, 20° left, 20° right).
Key findings
BlindPilot led participants to the chair significantly faster than the audio baseline overall (Wilcoxon signed-rank, p=0.04) and significantly faster on the Front-angle route (p=0.04); the Left (p=0.09) and Right (p=0.06) angles trended the same direction but did not reach significance. Reorientation events dropped from a mean of 2.9 (SD 1.3) per task on the audio system to effectively zero with BlindPilot (participants simply held the handle and walked). Five of six participants preferred BlindPilot in the post-interview (QUEST 2.0 adapted 7-point scale) across effectiveness, ease of use, feeling of security, and comfort; the one dissenter (P1) still acknowledged that the robot provided a greater feeling of security. Qualitatively, participants said the robot was 'similar to the feeling of walking with a person' (P5), contrasted with the audio system which 'required me to change my orientation repeatedly' (P2, P4, P6). Two design complaints were prominent: P1 and P3 wanted pre-movement context ('tell me where the destination is before you start'), and P1 and P6 criticised the robot's inability to let them control their own walking speed. Participants spontaneously named use cases beyond chairs: food-court seating (P2, P3), non-territorial offices (P3, P5), shop entrances (P1–P5), and train-door approach (P2, P4).
Relevance
For practitioners, this paper is worth reading for the 'local navigation' framing alone. Global and local navigation are not just different scales; they demand different interfaces. Audio turn-by-turn works reasonably well for macro routes where updates can be delivered every few seconds, but at the one-to-six-metre scale the latency of speech becomes disruptive and the user ends up walking a zig-zag. A physical guide — whether a leashed quadcopter, a suitcase robot, or a mobile platform with a handle — sidesteps this problem by making the approach trajectory the robot's problem, not the user's. Concrete design takeaways for future assistive-navigation systems: (1) carry the global-to-local handoff in the design, not as an afterthought; (2) provide pre-movement context ('I'm going to an empty chair, 6 metres ahead to the left') before physically starting to move; (3) build shared control over speed so the user can slow or stop without fighting the machine. Limitations are honest: only six participants, a controlled room with no crowd occlusion of the chair, no evaluation of how a busy food-court environment would affect landmark detection, and a bulky wheeled robot that is far from a take-home product. This is a CHI Late-Breaking Work proof-of-concept, not a deployment study.
Tags: local navigation · landmark object · blind navigation · indoor navigation · assistive robotics · visual impairment · object detection · orientation and mobility · shared control · LiDAR · late-breaking work