Dexbotic - The Force Spirit machine open source body intelligence VLA model one-stop research service platform

What's Dexbotic?

Dexbotic is an open source visual-linguistic-action (VLA) model one-stop research service platform for embodied intelligence by Dexmal, which solves the problems of fragmentation and low efficiency of embodied intelligence research. Based on PyTorch, Dexbotic provides a one-stop solution for research and development in the field of embodied intelligence.Dexbotic's core advantage lies in its unified modularized framework, which is compatible with a variety of mainstream large-language model interfaces, integrates embodied manipulation and navigation modules, and reserves the ability to expand, providing an architectural foundation for more complex task scenarios in the future. Dexbotic provides a high-performance pre-training base model, optimized for many mainstream algorithms, which significantly improves the performance in simulation and real tasks.Dexbotic supports integrated training in the cloud and locally, and adapts to a variety of research and development environments, providing a full chain of support for robot training and deployment, covering the complete process from simulation verification to the landing of the real robot.

Features of Dexbotic

• Unified Modular VLA FrameworkDexbotic is based on a visual language action model, compatible with mainstream large language model interfaces, integrated with body manipulation and navigation modules, and reserved for expansion capabilities, laying the architectural foundation for the future realization of complex mission scenarios such as full-body control.
• High-performance pre-trained base modelsThe system provides several pre-trained models for mainstream algorithms, such as Pi0 and CogACT, that perform better in simulation and real tasks, effectively reducing training costs.
• Experiment-oriented development frameworkIt adopts the system architecture of "hierarchical configuration + factory registration + entry distribution", which is in line with the principle of software opening and closing, and supports users to quickly switch tasks, models and parameters to realize efficient experiment iteration.
• Cloud and Local Integration TrainingIt is fully adapted to diverse R&D environments, supporting large-scale distributed training on cloud platforms such as AliCloud and Volcano Engine, as well as running efficiently on local machines equipped with consumer-grade graphics cards such as RTX 4090.
• All-link robot training and deploymentThe company provides standardized data interfaces and common deployment scripts for mainstream robotics platforms such as UR5, Franka, ALOHA, etc., covering the complete process from simulation validation to seamless landing of real robots.
• Open Source Hardware Support: Dexmal Force Spirit Machine has launched its first open source hardware product, Dexbotic Open Source - W1 (DOS-W1), to help research and development of embodied intelligence.
• Performance Improvement: Dexbotic's pre-trained model delivers significant performance gains across multiple mainstream simulators and real physical world tasks, for example, the Dexbotic version (DB-*) significantly outperforms the original model on all four tasks tested by SimplerEnv.

Dexbotic's core strengths

• Unified Modular VLA FrameworkIt takes the visual language action model as the core, is compatible with the mainstream large language model interface, integrates the body operation and navigation module, and reserves the expansion capability, which lays the architectural foundation for the realization of complex mission scenarios such as whole-body control in the future.
• High-performance pre-trained base modelsThe system provides several pre-trained models for mainstream algorithms, such as Pi0 and CogACT, that perform better in simulation and real tasks, effectively reducing training costs.
• Experiment-oriented development frameworkIt adopts the system architecture of "hierarchical configuration + factory registration + entry distribution", which is in line with the principle of software opening and closing, and supports users to quickly switch tasks, models and parameters to realize efficient experiment iteration.
• Cloud and Local Integration TrainingIt is fully adapted to diverse R&D environments, supporting large-scale distributed training on cloud platforms such as AliCloud and Volcano Engine, as well as running efficiently on local machines equipped with consumer-grade graphics cards such as RTX 4090.
• All-link robot training and deploymentThe company provides standardized data interfaces and common deployment scripts for mainstream robotics platforms such as UR5, Franka, ALOHA, etc., covering the complete process from simulation validation to seamless landing of real robots.

What is Dexbotic's official website?

• Project website:: https://dexbotic.com/.
• GitHub repository:: https://github.com/Dexmal/dexbotic.
• Hugging Face Model Library:: https://huggingface.co/collections/Dexmal/dexbotic-68f20493f6808a776bfc9fc4.
• Technical Report:: https://dexbotic.com/Dexbotic_Tech_Report.pdf.

Who Dexbotic is for

• Researchers in the field of embodied intelligence: Dexbotic provides a standardized, modular research infrastructure that enables researchers to rapidly conduct experiments, validate new algorithms and model architectures, and accelerate research progress in the field of embodied intelligence.
• Robotics Developer: For technicians working in robotics development, Dexbotic provides full link support from simulation to real robot deployment, and can be used to develop and optimize the vision, speech and motion control capabilities of robots.
• Artificial Intelligence Algorithm Engineer: Provides algorithm engineers with high-performance pre-trained models and a flexible experimentation framework to facilitate their algorithm optimization and innovation and explore new application scenarios.
• Students in higher education and research institutions: Dexbotic's open source nature and detailed documentation make it an ideal tool for students at universities and research institutions to study and research embodied intelligence, and to help train professionals in related fields.
• Enthusiasts interested in embodied intelligence: For enthusiasts interested in embodied intelligence but lacking a professional background, Dexbotic provides an accessible framework and a wealth of resources to help them get started and dive deeper into the field.