Octo<T>

Models & Types · AiDotNet.VisionLanguage.Robotics

Octo: open-source generalist robot policy trained on 800K demonstrations.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → Octo<T>

Implements: IVisionLanguageAction<T>

For Beginners

Octo is an open-source vision-language-action model for generalist robot control across different robot platforms. Default values follow the original paper settings.

How It Works

Octo (Berkeley, 2024) is an open-source generalist robot policy trained on 800K demonstrations from the Open X-Embodiment dataset. It uses a transformer backbone that processes interleaved image observations and language task specifications, outputting continuous robot actions that generalize across different robot embodiments and manipulation tasks.

References:

  • Paper: “Octo: An Open-Source Generalist Robot Policy (Berkeley, 2024)”

Example

// Create an Octo model for generalist robot policy
// open-source model trained on 800K robot demonstrations
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
            
// ONNX inference mode with pre-trained model
var model = new Octo<double>(architecture, "octo.onnx");
            
// Training mode with native layers
var trainModel = new Octo<double>(architecture, new OctoOptions());

Constructors

Constructor Summary
Octo(NeuralNetworkArchitecture<T> architecture, OctoOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
Octo(NeuralNetworkArchitecture<T> architecture, string modelPath, OctoOptions options)

Properties

Property Summary
int ActionDimension { get; }
int DecoderEmbeddingDim { get; }
int EmbeddingDimension { get; }
int MaxGenerationLength { get; }
string LanguageModelName { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> GenerateFromImage(Tensor<T> image, string prompt) Generates from image using Octo’s readout token + task-conditioned attention.
Tensor<T> PredictAction(Tensor<T> observation, string instruction) Predicts action using Octo’s readout token + diffusion head architecture.