What if you had to find your best friend in a crowd, but every single person was wearing the exact same robot suit?
For a robot trying to keep track of its teammates, this isn’t just a game of hide-and-seek—it’s a massive brain-teaser.
When robots look alike, they are “homogeneous agents”—which is like trying to tell the difference between two identical Legos while you are both running around.
Hafez
Farazi
The fact that our proposed network can learn to solve the highly complex problem of data association based solely on learning is promising.
The Homogeneity Problem
The Old Method
Usually, robots use math called a “Kalman filter”—a bit like a GPS that tries to guess where you’ll step next based on your last move. But these guesses often fail if a robot hides behind a wall or makes a sudden turn.
Building the Super-Brain
A “Memory-Gym” for Robots
Scientists have built a “super-brain” for robots using a 5-layer Deep LSTM network. Think of an LSTM as a “memory-gym” for a computer; it doesn’t just see what is happening now, it remembers what happened five seconds ago to predict the future.
The Real-World Test
Platform: igusr Humanoid Open Platform
The team used the igusr Humanoid Open Platform robots. These little metallic players have to recognize each other in real-time.
Giving Robots “Eyes” and a “Compass”
The Bouncer: Cascade of Rejectors
The researchers didn't just give the robots eyes; they gave them a “Cascade of rejectors”—which works like a high-speed bouncer at a club who only lets in the “correct” visual shapes.
The Compass: SVM for Heading Estimation
They also used an SVM to help the bots understand “Heading Estimation.” This is basically a digital compass that slices a circle into ten 36° classes to figure out which way a robot is facing.
Eye-Popping Results
The results were eye-popping. In a test of 3,140 frames, the new system had an identification success rate of 91.1%.
New System
91.1% success rate with the new Deep LSTM system.
Old Method
86.3% success rate with the old “Gold Standard” JPDA method.
Pinpoint Accuracy
Mean Localization Error: Only 0.22 m (about the length of a loaf of bread!). Older methods were off by as much as 0.67 m.
Blink-and-You-Miss-It Speed
Decision Time: 4 ms
It makes a decision in just 4 ms. That is way faster than you can blink!
The Reality Check
The Current Limitations
It isn’t perfect yet, though. The robots currently have to “whisper” to each other over Wi-Fi to share their directions. If the Wi-Fi drops, the robots might get confused.
The system also struggled when “false positives”—basically robot ghosts that aren't actually there—appeared right on top of real robots.
Key Takeaway: But for now, these bots are proving that a good memory is the secret to never losing a friend in a crowd.
Reference: “Online Visual Robot Tracking and Identification using Deep LSTM Networks” — Hafez Farazi and Sven Behnke (Autonomous Intelligent Systems Group, University of Bonn).