A GPU-accelerated simulation environment for training generalist autonomous agents.

Python, pytorch

CURRENT ENVIRONMENTS: FishTank - A 2d procedurally generated fish tank containing food pellets and a physics simulation driven using Verlet integration. Each character in the environment receives a sequence of observations generated by casting rays out from the character's head. This 1D RGB image is flattened into a single array, and appended with the character's current actions, their body position as a percentage of the tank size in the x and y dimension, their direction in terms of normalised x and y coordinates, and their velocity is normalised by their maximum acceleration.

The controller then processes the input and returns an array of outputs Actions correspond to [forward_thrust (0 - 1), rotational acceleration (0 full counter-clockwise, 0.5 - still, 1 - full clockwise, (TBD)]

CONTROLLERS: The primary controller AgentController, is built using a transformer architecture which predicts the future observation state t+1 at time t. The architecture uses a context length of 1024 samples, 12 layers and 12 heads, and an embedding size of 764 units. Models can be dynamically saved and loaded from the checkpoints folder, using an existing model config if found.