FluxVLA Engine: A One-Stop VLA Engineering Platform for Embodied Intelligence

This release expands FluxVLA with ARM reward modeling, GR00T RoboCasa training/evaluation support, a unified dataset pipeline, and Triton-accelerated Pi0.5-RTC inference.

Highlights

• Add ARM reward modeling with RA-BC/AW-BC reweighting, including ARM dataset/model components, weighting utilities, progress inference scripts, and setup documentation.
• Add GR00T RoboCasa support with a new training config, RoboCasa evaluation runner, RoboCasa-specific transforms, checkpoint utilities, and simulator dependency checks.
• Add Pi0.5-RTC Triton kernel inference acceleration for lower-latency real-time control.
• Unify the dataset pipeline and expand multi-policy VLA support, including updated OpenVLA LIBERO LoRA configs and shared preprocessing utilities.
• Update README and task documentation for ARM, RoboCasa, SARM, and inference acceleration workflows.

Upgrade Notes

• RoboCasa support now requires the patched robosuite dependency and compatible mujoco version specified in requirements.txt.
• The older RLDS dataset path has been removed in favor of the unified dataset pipeline.
• Some OpenVLA full-train configs were replaced by LIBERO LoRA finetuning configs.