MLAgents - 2D Bike game AI Agents not improving

I am working on 2d Bike racing game. My main problem is that i am able to beat AI player via manual playing.
It's like CPU vs User game.

ON observation i am using Checkpoint position - bike position
and using using Velocity vector

Adding positive rewards on each checkpoint crossing + end episode.
Subtracing rewards on each Stop action (For example i am using forward and backward <Onbackward> i am giving negative reward)
Subtracting for each tilt condition

Using DiscreteActions : 2
I have been training my model from 36 hs on AWS instance

 

Paid Help needed please revert