mbaske
fundamental Questions and compatibility in year 2024
As there is not any update for newer mlagents and python version, I decided to open this issue and ask some questions. As I tried to train my environment in ML- agents by rays and camera too, I found mbaske's 3D grid sensor. But unfortunately I got some error to handle it with ML-agent 3.0.0. Finally I found that I should downgrade my python(now 3.6) and mlagnet(2.21-exp.1). so I have still some question about it that maybe is helpful for everyone who are willing to work with that. it stores the agent's position local or global? I mean when I put the 3D sensor on my agent it stores the position as local?
Observations are always local, so that they're independent from where agents are in the environment. The sensor tracks positions of observable objects relative to the sensor/agent position.
thanks for your response, the 3D grid sensor senses(or collect) all info inside it too? or like raycasts it gets just the info whatever it hits? for example suppose I defined a sphere (Lat Angle North and Lat angle south = 90 and Lon angle = 180), it gets the information from all collider inside this sphere too? or just what is on the surface?
The sensor detects sets of points that need to be generated upfront. That's why detectable gameobjects need to have the corresponding class, please see https://github.com/mbaske/grid-sensor?tab=readme-ov-file#Detecting-Game-Objects The ml-agents toolkit comes with its own grid sensor. Last time I checked, their sensor worked by physics overlap checks against colliders, which can drag down performance quite a bit. My sensor is lighter on the physics engine, but the trade-off is that more setup work is required.