LEAPP#

nvidia-isaac/leapp

Lightweight Export Annotations for Policy Pipelines

Export the whole policy pipeline, not just the model weights.

LEAPP helps teams move learned policies and multi-stage PyTorch pipelines from research code into deployable artifacts. It traces real execution across preprocessing, model inference, postprocessing, constants, and tensor library glue, then exports per-node models and a YAML pipeline specification that downstream runtimes can consume.

LEAPP is designed for the full policy path, including the logic that surrounds the learned model: normalization, feature transforms, recurrent state, action chunking, unit conversion, clipping, and task-specific output shaping. Without LEAPP, teams often rewrite that logic in another language, maintain a separate export pipeline, or carry the full research environment into deployment.

With LEAPP, you mark the input and output boundaries of the compute you already run. Execute the annotated pipeline with sample inputs, and LEAPP compiles the traced graph into portable artifacts.

Get started

Read the guides

Is LEAPP a Fit?#

Use LEAPP when your deployment target needs more than a single model checkpoint. It is a strong fit when you need to:

Deploy Python policy code

Move Python-based policy code from research workflows into deployment.

Capture the full pipeline

Keep preprocessing, postprocessing, constants, and tensor operations around the learned model.

Avoid specialized export-only implementations

Eliminate the need for a separate export-only implementation of the same computation.

Move between runtimes

Carry policy logic into another runtime, simulator, application, or deployment stack.

Export modular policy graphs

Represent a policy as connected nodes, so each stage can be inspected, exported, and wired independently.

Describe tensor meaning

Help downstream systems connect inputs, state, commands, and outputs.

How LEAPP Fits In#

LEAPP traces the code path you actually execute. It records tensor flow between annotated stages, exports each stage independently, and writes a YAML specification describing how the exported pieces connect at inference time.

The result is an export bundle that can be inspected, validated, and handed to downstream systems without requiring those systems to understand the original training project.

annotate inputs
Existing policy code
annotate outputs
export bundle
node1.pt node1.onnx pipeline.yaml visualization.png

Get Started#