OpenDriveLab/AgiBot-World

Features

• Bimanual Manipulation Training - Develops robotic models for complex bimanual manipulation using foundation models and large trajectory datasets.
• Embodiment Adaptations - Implements a suite for adapting pre-trained models to different robot bodies without requiring new multi-embodiment training data.
• Cross-Embodiment Transfer - Features a suite that adapts pre-trained models to different robot bodies without requiring new multi-embodiment training data.
• Robotic Trajectory Standardizers - Ships utilities to transform raw robotic data from various formats into standardized layouts for training.
• Demonstration Tracking - Implements training loops that utilize human-reference trajectory data to guide the learning of complex physical tasks.
• VLA Model Validation - Provides a framework for validating vision-language-action models through open-loop testing and physical hardware replays.
• Robotic Policy Evaluators - Implements a platform for measuring accuracy and generalization in robotic manipulation tasks using large scale datasets.
• Open-Loop Evaluations - Includes a pipeline to replay recorded actions and measure policy accuracy before real-world hardware execution.
• Physical AI Fine-Tuning - Refines autonomous driving and manipulation models using real-world data and production-validated infrastructure.
• Trajectory Standardization - Implements a unified data format that normalizes diverse robotic sensor streams and actions for consistent training.
• Scalable Robot Policy Trainings - Provides a comprehensive infrastructure for developing bimanual manipulation policies using foundation models and human-reference data.
• Bimanual Manipulation Trainings - Develops bimanual manipulation policies using large-scale trajectory datasets and pre-trained foundation models.
• Robotic Arm Training - Uses human references and adaptive tracking to enable robotic arms to perform complex physical manipulation tasks.
• VLA - Provides a testing suite for vision-language-action models using open-loop validation and hardware replays.
• Robot Learning Platforms - Offers a platform for measuring accuracy and generalization in robot manipulation tasks using large-scale trajectory datasets.
• Custom Data Fine-Tunings - Adapts pre-trained robotic policies to specific tasks or new environments using custom user-provided datasets.
• Standardized Evaluation Protocols - Defines a protocol for controlling environmental noise and lighting to ensure reproducible robot testing results.
• Inference Execution Models - Executes trained policies via local controllers or remote servers to translate sensor observations into robotic actions.
• Adversarial Robustness Training - Provides methods for introducing synthetic noise and environmental shifts to improve model robustness during training.
• Training Data Generation - Provides processes for introducing environmental changes and task shifts to increase training data efficiency.
• Autonomous Driving Refinements - Provides a production-validated environment for fine-tuning autonomous driving models using real-world data.
• Modular Physical Architectures - Ships a lightweight physical architecture with interchangeable components for multimodal grasping and various body configurations.
• Dexterous Robot Hand Design - Provides a modular, lightweight hardware design enabling multimodal robotic grasping and manipulation.
• Robotics Visualization Tools - Provides tools for rendering camera streams, robot states, and actions from trajectory datasets to inspect data quality.
• Policy Servers - Provides a decoupled execution model that hosts policies on GPU servers and streams actions to local controllers.
• Execution Fidelity Validation - Provides a method for replaying recorded demonstrations on hardware to measure execution fidelity against original recordings.

Star history