Unity ML Agent fails to learn simple endless runner

I’m playing around with Unity ML and try to train an agent to play a simple endless runner.

So far the agent fails to learn from the first obstacle and continues to jumps into it.

As observations, I use the distance to the next obstacle for now. Actions are 0 - do nothing; 1 - jump. The agent gets rewards for the time alive.

Source code here:

Current training config:

gamma: 0.99
beta: 1e-3
epsilon: 0.2
learning_rate: 1e-5
buffer_size: 40960
batch_size: 32
time_horizon: 2048
max_steps: 5.0e6

Any ideas or hints on how to enable convergence? How to make it learn?


Same problem right here. We built a runner game and trying to implement the AI stuff in playing this game itself. But so far the agent seems to be good at nothing but committing suicide. We should never consider doing this from the start as our final project, it is due tomorrow, and we don’t know what to do. ,You know what, we’re having the same problem. We build a game and trying to implement it into AI surrounding for our final project. But what the outcome just turned out as exactly what you said. Our agent is good at nothing but committing suicide, and it seems that it is particularly good at jumping into pitfalls. We have no idea what to do. Tomorrow is the due of our final project. We should never consider doing this from the start.