avalanche
A swarm of simulated snowboarders shares one policy and learns one mountain, trained with PPO on a CPU.
1,055 mmean distance down the run after 3.2 minutes of CPU training, up from 21 m
The riders are batched physics trained with PPO on the CPU and watched live in the browser. In a smoke run, mean distance went from 21 m to 1,055 m in 3.2 minutes: 15.7M samples at about 79k samples per second. A resumed live run reached a 26.6 percent finish rate at 275M samples.
Flips show up early in training and fade as distance climbs. Why is not established. There is no recorded video yet, only the live viewer.