Summary
Embodied.cpp is a portable C++ inference runtime for deploying VLA and World-Action Model (WAM) models on heterogeneous edge hardware (CPU/CUDA GPU/NPU) using GGUF weights. It organizes the shared execution path of embodied models into five layers — input adapters, sequence builders, backbone execution, head plugins, and deployment adapters — enabling multi-rate closed-loop control, latency-first batch-1 inference, and extensible embodied I/O across different robots and simulators from one backend.
Key Contributions
- Unified five-layer architecture that captures the shared execution path across both VLA and WAM models, enabling one backend to deploy both model families
- Latency-first batch-1 inference design specifically for closed-loop embodied deployment, unlike existing runtimes built for request-response serving
- Demonstrated 100% and 91% closed-loop task success rates for VLA deployments; WAM benchmark reduces block memory from 312.2 MiB to 88.1 MiB
Significance
Solves the fragmented Python-stack deployment problem for production robot embodied AI, making VLA and WAM models practically deployable on constrained edge hardware without model-specific glue code.