ARCMove SoftwareDevelopment Update.
Where we are.
v0 prototype
Person Recognition integration
v0.5 · GTM ready
v1.0 launch
Detect.Then forget.
Front camera only. Once the user steps out of the frame, the robot forgets who they were — no memory of the person it was just with.
User in frame.
User out of frame.
User returns — but as a stranger.
Detection is not recognition. And anything behind the robot is invisible.
Improve the user experience while navigating them to the goal.
The robot should remember who's using it — and have an all-around view.
Same user.Both cameras.
Front + back. The robot remembers who it's leading — through the turn, into the follow.
How itremembers.
Two ML models. A body fingerprint and a face fingerprint. Body first, face confirms.
Find the person.
A vision model spots people in the frame and crops out the body and face for the next stage.
Two fingerprints.
ML turns the body and face into mathematical signatures — vectors a computer can compare in microseconds.
Structure, then face.
Body match scans the memory bank; the face gate confirms it's the same person — even after the user turns and walks away.
See itin action.
Happy-path run: interaction on the front camera, the user moves to the back, robot keeps the same identity.
On the robot.Real view.
Cameras mounted in their actual positions on arcmove — same identity tracked across the real front→back perspective the robot sees in deployment.
Robot in motion.Lock holds.
Cameras moving with the robot — testing the identity lock under real navigation motion: turning, driving forward, the user shifting view as the robot moves.
Where itfalls short.
A working POC, not a finished product.
Three honest limits to know about.
Side blind spots.
Only front and back cameras. A user crossing left or right disappears between the cones.
Fix: 4 cameras in iteration 3.
Face when visible.
Face confirms identity — but only up close. Far away or facing away, only the body is used.
Mitigation: body bank stores multiple poses.
One user at a time.
Once selected, the robot locks to that user only. No parallel tracking.
By design: one guest per interaction.
All-around.No blind spots.
From 2 cameras to 4 — front, back, left, right — with overlapping FoVs so the user never leaves coverage.
Two more cameras.
Left and right cameras added to the existing front + back. FoVs are sized to overlap at the corners — no gap between cones.
Port to Jetson Orin Nano.
Move the perception stack off the dev machine onto the robot's small on-board Jetson — efficient inference, lower BOM cost, no laptop tethered to the robot.
Wire into navigation.
Feed the identity signal into the robot's navigation and behaviour stack — turning "we know who's there" into "the robot does what we want."
Thank you.
Questions & discussion.