This repository is a comprehensive research codebase for Vision-Language-Action (VLA) models in robotic manipulation. It implements multiple model architectures (PI0, PI05, GR00T N1.5/N1.7, Cosmos3, FastWAM, DreamZero, DiT4DiT, SmolVLA, OpenVLA, LLaVA) with flow matching, diffusion, and autoregressive action heads. The primary purpose is to train, evaluate, and deploy VLA policies, with full pipelines for data ingestion (LeRobot parquet datasets), transformation (normalization, augmentation, video decoding), training (FSDP, mixed precision, gradient checkpointing), and inference (simulation benchmarks and real-robot ROS-based runners). Additionally, it includes research contributions in reward modeling (ARM, SARM, RA-BC/AW-BC), progress reconstruction, and subtask annotation. The codebase is highly configurable via mmengine-style configs and a registry system. While predominantly research-oriented, it also serves as an application framework for real-robot deployment with CUDA and Triton kernel optimizations, ZMQ-based serving, and Jetson Orin compatibility.