UInferenceInterfaceOnnx is a concrete implementation of the UInferenceInterface using ONNX Runtime. It loads .onnx files and performs inference given a vector of float observations, returning a serialized action string.

This class serves as the default backend for model inference in the framework.

ONNX (Open Neural Network Exchange) is a widely used standard format for trained neural networks. This implementation wraps ONNX Runtime and plugs directly into the RL bridge infrastructure.

Key features:

• Loads .onnx files from disk
• Runs inference on CPU
• Designed for 1D float observations with shape [1, N]
• Outputs flat float arrays as comma-separated strings

Instantiate UInferenceInterfaceOnnx and pass it to a bridge or component via SetInferenceInterface().

After calling LoadModel(), you may begin using RunInference() with live observations.

Initializes ONNX Runtime if not already active, then creates a session from the given model path.

• Model must have input name "obs" and output name "actions"
• Path can be absolute or project-relative
• Returns true if the session loads successfully

Performs inference using the loaded session:

• Wraps input array into ONNX tensor format
• Runs the model on CPU
• Reads result from "actions" output tensor
• Converts output array into a comma-separated FString

EXPECTED FORMAT: Input array must match the model's observation space. For example:

TArray<float> Observation = { 1.0f, 0.0f, -0.5f, 3.2f };

This array will be treated as a [1, 4] tensor and passed to ONNX Runtime.

The output will be formatted with a number of values equal to your policies action space. For example, if your action space has a size of 2:

"0.1,0.9"

Returns true if the ONNX session is valid and ready.

• Uses Ort::Env, Ort::Session, and Ort::SessionOptions to manage runtime state
• Globals GEnv and GSessionOptions are shared across instances
• Inference expects a 2D shape [1, N] where N = Observation.Num()
• Output must be cast to float array and reshaped manually