Question Need help with ML Agent avoid obstacle

Hi everyone!

I have problem with my agent, I tried to train them to avoid obstacle and collect the ball. I already train with 50,000,000 max step but it still not work.
Here is my logic to train agent:
- OnEpisodeBegin: I just reset agent's position.
- CollectObservations:
+ Collect ball postion.
+ Collect agent position.
+ Collect direction from agent to ball.
+ Collect distance from agent to ball.
+ Collect environment info via RayPerceptionSensor3D component.
- OnActionReceived:
+ Just move and rotate agent with direction get from actions.ContinuousAction.
- Reward:
+ +5 when agent got ball.
+ -2 when agent hit the wall or obstacle.
After agent get ball or hit wall I also call EndEpisode.

My config:

behaviors:
CharAgent:
trainer_type: ppo
max_steps: 50000000
time_horizon: 64
summary_freq: 10000
keep_checkpoints: 5
hyperparameters:
batch_size: 512
buffer_size: 2048
learning_rate: 0.0003
beta: 0.05
epsilon: 0.2
lambd: 0.99
num_epoch: 3
learning_rate_schedule: linear
network_settings:
normalize: false
hidden_units: 128
num_layers: 2
vis_encode_type: simple
reward_signals:
extrinsic:
gamma: 0.99
strength: 1.0

**** As I think right now is the agent auto learn to go around the obstacle or did I missing something to let the agent know how to move around the obstacle?

Result after train: Agent still can't move around the obstacle to get the ball.
The image below show the scene I setup to train agent.


Anyone have idea about my problem, I just play with MLAgent for a week.
Please give me advise.
Thank you.