Playing tag with robots
Real-time tracking under real-life constraints - by Anastasios Vlachos
Imagine two robots playing tag, i.e. a follower trying to catch a leader, where the leader robot moves freely, and the follower, tries to follow it closely.
The catch is that the follower doesn’t know the leader’s path in advance. The best it can do is to try to predict where the leader will go - based on the leader’s movements so far - and adjust its own movements accordingly.
This process of predicting and adjusting on-the-go is called Online Tracking:
At each timestep, the follower observes the leader’s position, predicts where it will go next using only the information gathered up to that point, and moves based on that prediction.
To evaluate how well the follower tracked the leader after the game, we use the concept of Regret. Regret essentially measures how much better the follower could have tracked the leader if it had known its entire trajectory in advance - where it would have performed perfect tracking. The follower experiences regret when it finds itself far from the leader.
Now, imagine scenarios where the follower has to track a more agile leader, loses sight of the leader because something blocks its view, or receives corrupted information about the leader’s position!
We want to understand how these challenges impact the follower’s ability to track the leader - that is, how much additional Regret is incurred due to these complexities. In simple terms, we want to teach robots to track moving targets as effectively as possible under real-world conditions and quantifying how much performance is lost due to these imperfections.
Text by Anastasios Vlachos; image created with Microsoft Designer
Real-time tracking under real-life constraints - by Anastasis Vlachos