Lamp: Unified Compiler for Edge‑Deployed Embodied AI, Bridging Robot Prototypes to Products

The Meet AI Compiler technical salon highlighted that existing AI compilers for autonomous driving or internet chips do not directly apply to embodied intelligence because the algorithm structures lack a unified paradigm and face strict resource and data constraints on robot edge devices.

Three common solution routes are described: (1) Vision‑Language‑Action (VLA) that combines VLM, LLM and an Action head, but suffers from limited compute resources and insufficient training data; (2) a brain‑cerebellum split where a large model handles planning and a smaller module handles execution, similar to agentic solutions; (3) a deterministic workflow that stitches together AI and non‑AI atomic capabilities into a fixed pipeline.

These diverse routes create a challenge for AI compilers: they must handle a constantly evolving set of pipelines rather than a single static model. Consequently, the compilation target shifts from model‑level optimization to full pipeline compilation, requiring support for multiple modules, glue layers, and rapid productization.

Lam p is introduced as a unified compiler framework designed for the edge deployment of embodied AI. It provides five key capabilities:

End‑to‑end DAG management : captures the entire sensor‑to‑action pipeline as a graph, including data types, shapes, device, quantization, and state information.

Grouped compilation with multi‑backend coordination : separates horizontal dimensions (hardware platforms such as NVIDIA GPUs, Thor, Orin, domestic chips) and vertical dimensions (kernels from hand‑written CUDA, CuTeDSL, CUTLASS, vendor‑specific kernels).

Unified runtime execution : loads compiled artifacts, reads ABI and manifest, performs weight binding, resource allocation, kernel loading, and scheduling.

Full‑link closed‑loop delivery : versioned artifacts are packaged with verification evidence, enabling traceable deployment, regression testing, and performance profiling.

White‑box controllability and deep tunability : retains intermediate graph states and profiling data, allowing rapid pinpointing of correctness or performance regressions down to specific kernels or resource allocations.

In practice, Lamp reduces two major costs: manual rewrite effort (by converting Python prototypes directly into a unified DAG, avoiding extensive C++ porting) and regression effort (by locating changes in the graph rather than re‑checking the whole pipeline). The framework also supports dynamic tracing, fusion, and translation of pipeline components into groups or scopes before applying traditional compiler optimizations such as TensorRT or TVM.

Performance results show that the original π‑0.5 inference frequency of 1.4 Hz was increased to 22 Hz on real hardware, approaching real‑time "cerebellum" operation. Additional optimizations using Triton raised performance to ~3.5 Hz, while graph‑level compilation achieved ~8 Hz with better cost‑effectiveness.

Overall, Lamp enables complex robot algorithms to move from flexible Python prototypes to production‑grade edge deployments with automated build, verification, and continuous optimization, while maintaining transparency and white‑box insight into the entire compilation and execution flow.