open-gigaai/giga-brain-0

Features

• Observation-to-Action Mappings - Translates high-dimensional sensor data into specific robot control signals or high-level task subgoals for physical execution.
• Robot Action Model Frameworks - Provides a comprehensive framework for training and deploying neural networks that map multi-modal sensor data to physical robot control signals.
• Embodiment Adaptations - Adapts pre-trained vision-language-action models to specific robot hardware configurations using specialized datasets and fine-tuning.
• Robot Embodiment Fine-Tunings - Provides a specialized pipeline to adapt pre-trained base models to specific robot hardware configurations and sensor sets.
• Model Training Pipelines - Provides end-to-end workflows for pre-training and fine-tuning action models on specific robot embodiments.
• Scalable Robot Policy Trainings - Trains neural networks to predict robot movements and subgoals across diverse physical embodiments.
• Robotic Manipulation Models - Executes dexterous, long-horizon physical tasks by mapping visual and language inputs to robot arm control actions.
• Robot Policy Inference - Executes pre-trained policies on real-time robot sensor data to predict immediate control actions.
• Model Fine-Tuning - Adapts pre-trained vision-language-action models to specific robot hardware configurations using specialized datasets and fine-tuning.
• Language and Robotics Integration - Integrates natural language understanding with physical robot control to execute complex, long-horizon manipulation tasks.
• Robot Action Streaming Servers - Provides a server architecture to stream real-time control signals and inference results to robotic hardware.
• Policy Servers - Implements a GPU-backed server that receives multi-modal observations and returns real-time control actions via a network protocol.
• Remote Inference Streaming - Implements a persistent network connection for streaming sensor observations and receiving real-time control commands from a remote model.
• Feature Scale Normalization - Standardizes raw sensor inputs and actuator outputs into a uniform scale using statistical scaling for stable model convergence.
• Training Pipelines - Provides structured workflows that decouple model architecture from training logic to support both scratch training and fine-tuning.
• Multi-modal Embedding Generation - Combines visual and linguistic data into a shared vector representation to generate continuous or discrete robot actions.
• Multimodal Action Prediction - Generates continuous or discrete robot control commands based on combined visual and linguistic inputs.
• Robotic Data Processors - Converts raw sensor data into uniform formats and calculates normalization statistics for robot states and actions.
• Training Pipelines - Provides structured workflows that decouple model architecture from training logic to support both scratch training and fine-tuning.
• Embodied Foundation Models - World model-powered vision-language-action framework.

Star history