huggingface/lerobot

Features

• Expert Imitation Learning - Trains transformer-based models on demonstration datasets to enable autonomous robotic manipulation.
• Edge AI Model Deployment - Executes optimized machine learning models on physical robotic hardware for autonomous tasks.
• Robot Learning Platforms - Provides a comprehensive research platform for the end-to-end lifecycle of robotic learning.
• Robotics Middleware - Provides a unified interface to decouple high-level control logic from specific robotic hardware implementations.
• Hardware Abstraction Layers - Decouples high-level control logic from specific communication protocols of diverse robotic hardware.
• Machine Learning Training - Provides standardized pipelines for training machine learning models for robotic manipulation tasks.
• Interaction Data Collectors - Records sensor inputs and motor commands into standardized formats for training datasets.
• Robotic Policy Evaluators - Executes standardized benchmarks to test trained policies in simulated and physical environments.
• Real-time Policy Execution - Executes trained machine learning models on physical robots by mapping outputs to motor control signals in real time.
• Robotics And Autonomous Systems - Maps trained machine learning model outputs to motor control signals for real-time autonomous robotic behavior.
• Kinematics - Calculates forward and inverse kinematics to translate between joint space and Cartesian coordinate systems.
• Robotics and Control - Optimizes neural network models using interaction data to enable autonomous behavior on physical robots.
• Robotics Libraries - Provides a modular library for training imitation and reinforcement learning models for physical robots.
• Benchmarking Suites - Runs standardized rollout loops across tasks to aggregate performance metrics like success rates.
• Distributed Training - Distributes training workloads across multiple GPUs to accelerate robotic policy learning.
• Distributed Training Orchestrators - Parallelizes the optimization of complex robotic policies across multiple graphics processing units.
• Inference Rollout Evaluation - Runs trained models on physical hardware to perform inference and record episodes for performance assessment.
• Dataset Management Tools - Standardizes data storage using synchronized video and state files for efficient dataset management.
• Remote Inference Streaming - Offloads computational workloads to external servers while maintaining real-time control.
• Remote Inference Streaming - Streams robot observations to remote servers for real-time policy inference on resource-constrained hardware.
• Modular Policy Architectures - Implements imitation, reinforcement, and vision-language-action models in a modular structure.
• Policy Architectures - Organizes neural network components into interchangeable blocks for varied learning models.
• Foundation Models - Community library for end-to-end robotic learning and VLA deployment.
• Machine Learning - Pretrained models and datasets for real-world robotics.
• Robotics Foundation Models - Efficient vision-language-action model trained on community data.
• Robotic Platforms - State-of-the-art AI framework for real-world robotics.
• Robotics Libraries - PyTorch-based models and tools for real-world robotics.
• Action Normalizers - Scales raw action values into a normalized range for model training and inference.
• Robotic Interaction Schemas - Structures interaction data into synchronized video and state files for cross-environment compatibility.
• Performance Benchmarking - Runs standardized evaluation loops and rollout tests to measure robotic policy performance.
• Action Coordinate Transformers - Converts action data between coordinate systems to support different training requirements.
• Observation Processors - Applies environment-specific transformations to observation data to handle unique sensor formats.
• Observation Standardizers - Maps simulator-specific outputs to a unified key convention for policy compatibility.
• Normalization Migrators - Extracts normalization layers from trained model weights into external processor pipelines.
• Legacy Dataset Replayers - Applies coordinate transformations to historical data to ensure compatibility with updated hardware.
• Simulator Wrappers - Wraps third-party simulators into a standardized interface for robotic policy training and evaluation.
• Robotic Data Processors - Provides external processors to decouple input and output scaling from core model weights.

Star history