What You Are Seeing
Start the patrol to let LOCI build memory.
The robot stores each observation as a memory that keeps the visual embedding together with where and when it happened.
Current view: waiting for live observations.
Next best step: click Build Memory, then let the robot complete part of one aisle.
Map Legend
Robot
Selected area
Matching memories
Similarity anchor
Predicted route
Surprise obstacle
Upcoming patrol path
Start here: build memory first. Once LOCI has enough memories, the guide buttons will tee up the best interactions automatically.
Guided Demo
Record a clean walkthrough in five clicks
Use the guide to stage the strongest moments automatically, then use the manual explorer below if you want to dig deeper.
Idle
Waiting for the robot to start building memory.
Stored memories
0
Prediction
warming up
Next waypoint
(2, 18)
Route progress
0%
Start the patrol and let LOCI watch one aisle fill up before moving into search and surprise detection.
Where + when + what
Each memory keeps the observation together with its location and time window.
Skip irrelevant history
LOCI narrows the search before ranking, instead of scanning every stored vector.
Explain surprises
Prediction plus retrieval shows when the robot is entering something unfamiliar.
Manual Explorer
What happened in this area?
Drag a box over the map to ask LOCI for memories from a place and time window.
Narrator cue: use this after the robot has finished one stretch of the aisle, so the result feels like browsing a real memory.
to
Why this is fast
Traditional vector databases scan every embedding with independent range filters. LOCI narrows space with Hilbert curve bucketing and skips irrelevant time windows with epoch sharding before ranking memories.What looked like this before?
Click a point on the map to find past moments that look similar nearby.
Narrator cue: this is the easiest way to explain that LOCI remembers both appearance and physical context.
Why this matters
LOCI combines vector similarity with spatial proximity. A normal vector DB can rank similar embeddings, but it does not naturally explain what was similar here inside a physical region and time window.Is something unexpected ahead?
Ask LOCI what the robot expects a few steps ahead, then compare that prediction with memory.
Narrator cue: add a surprise obstacle first, let the robot get close, then run this step to show the surprise score rise.