dexmal/dexbotic

What it solves

Dexbotic addresses the complexity of developing Vision-Language-Action (VLA) models for embodied intelligence. It provides a unified toolbox to streamline the entire lifecycle of VLA development—including pretraining, fine-tuning, inference, and evaluation—reducing the need for fragmented setups when working with different robotic policies.

How it works

Built on PyTorch, the toolbox uses a modular architecture based on layered configurations, factory registration, and entry dispatch. This allows researchers to swap models, modify task configurations, or add new capabilities via experimental scripts without rewriting core logic. It supports a unified training data format for various robots and provides optimized pre-trained foundation models for mainstream algorithms. The system is designed to work across local consumer GPUs and cloud platforms, with specific support for high-performance hardware like Blackwell GPUs.

Who it’s for

It is designed for researchers and developers in the field of embodied intelligence who are building or evaluating VLA models for robotic manipulation and navigation.

Highlights

• Broad Model Support: Integrates mainstream VLA policies including π0, CogACT, OFT, MemVLA, and DM0.
• Robot Compatibility: Provides unified data formats and deployment scripts for robots such as UR5, Franka, and ALOHA.
• Flexible Training: Supports both local and cloud training, including FSDP2 for distributed training and GRPO for reinforcement learning.
• Optimized Inference: Includes a unified v1 inference API and a Triton-backed realtime backend for significant speedups.
• Comprehensive Benchmarking: Ready-to-use evaluation pipelines for environments like Libero, CALVIN, SimplerEnv, ManiSkill2, and RoboTwin2.0.