starVLA/starVLA

What it solves

StarVLA is a modular research platform designed to simplify the development of Vision-Language-Action (VLA) models for generalist robots. It addresses the complexity of building these systems by providing a "Lego-like" codebase where components like model backbones, action heads, and training recipes can be swapped and integrated without rewriting the entire pipeline.

How it works

The platform uses a decoupled, composable stack that separates the training infrastructure from the specific model architecture. It allows users to plug in different foundation-model backbones (such as Qwen-series or Florence-2) and interchangeable action heads (e.g., autoregressive discrete tokens, parallel continuous decoding, or flow-matching experts). The system maintains a standardized model-forwarding interface and a model-agnostic dataloader, ensuring that data flows consistently regardless of the chosen framework variant.

Who it’s for

It is primarily intended for robotics researchers and developers who are building generalist robot policies and experimenting with different VLA architectures, training paradigms, and benchmarks.

Highlights

• Modular Architecture: Supports rapid prototyping by allowing independent debugging and "smoke testing" of submodules.
• Diverse Action Heads: Includes implementations for StarVLA-FAST, StarVLA-OFT, StarVLA-PI, and StarVLA-GR00T.
• Broad Benchmark Support: Integrated with numerous benchmarks including SimplerEnv, LIBERO, Robocasa, RoboTwin 2.0, and RoboDojo.
• Flexible Training: Supports supervised fine-tuning (SFT), multimodal co-training, and RL post-training via RLinf.
• Hardware Compatibility: Supports training on Ascend NPUs and single A100 GPUs for smaller VLMs.