- Inference acceleration for PI0.5 and Gr00t (CUDA graph, CUDA and triton ops)
    - Multi-policy support: PI0, PI0.5, GR00T, OpenVLA and LlavaVLA
    - CUDA graph optimization and RTC acceleration for real-time robot control
    - Embodiment configs for ALOHA, Tron2, Libero, and UR3 evaluation pipelines
    - Safetensors support for 3x model size reduction and fast checkpoint loading
    - Multilingual documentation (EN / ZH / JA)