New Updates on My Robotics Work

New Updates on My Robotics Work#

Yesterday, I devoted my attention to designing controllers for LoTIS.

LoTIS takes two inputs: a reference trajectory, recorded simply as a sequence of RGB images, and the drone’s current camera view. For every frame of that trajectory, it predicts three values: the image-space coordinates where that point would appear in the current view, whether the point is visible at all, and a normalized distance to it. Because it processes the entire trajectory jointly rather than selecting a single subgoal, it remains reliable even when the drone starts some distance from the route. It also requires no camera calibration, pose data, or robot-specific training, which makes it independent of the platform it guides.

That independence comes at a price: LoTIS deliberately separates perception from action, so it says where the path lies but not how to fly there. A controller is therefore required to translate these three values into actual flight commands.

I explored both a reinforcement learning approach and a deterministic, rule-based one. The results were decisive: the reinforcement learning controller performed substantially better.

It is a fitting reminder of how much learned methods now offer. Here’s to the age of AI!