Building an RL Traffic Agent for Indian Roads, Part 8: The Ramp That Wouldn't Go AwayPart 7 ended on a table that looked clean. Nav scaling fixed the steering lock-in, extending training to 100,000 steps gave the CNN room to actually learn road geometry, and every metric in that closiAug 8 · 10 min readlocal-minima.hashnode.dev
Building an RL Traffic Agent for Indian Roads, Part 7: Unlocked, and Immediately Falling Apart DifferentlyThe critic had curvature. The actor finally had somewhere to go. Which meant, for the first time, the car actually moved. It also meant a new set of problems, because "moving" and "driving well" turneAug 8 · 7 min readlocal-minima.hashnode.dev
Building an RL Traffic Agent for Indian Roads, Part 6: The Frozen RobotEverything in Part 5 worked. The pipeline ran end to end, LIDAR in, occupancy grid out, latent vector through the actor, continuous steering and throttle out the other side. Then I actually let it tra1Aug 8 · 13 min readlocal-minima.hashnode.dev
Building an RL Traffic Agent for Indian Roads, Part 5: The Math, the Bugs, and Everything That BrokePart 4 was the reasoning that got me to an architecture on paper. This part is what happened once I actually tried to build it, which turned out to be a very different exercise. Most of Part 4 was "whAug 8 · 18 min readlocal-minima.hashnode.dev
Building an RL Traffic Agent for Indian Roads, Part 4: The Architecture After going through the theory, I didn't want to just jump into code. I wanted a base to build on first, a shape for the whole thing before I touched a single line of Python. So the first question I hAug 8 · 19 min readlocal-minima.hashnode.dev
