LEAPP: Lightweight Export Annotations for Policy Pipelines

Full Isaac LEAPP docs

Export the policy code you already run: preprocessing, model inference, and action processing together.

Isaac LEAPP is NVIDIA's library for machine learning pipeline export. It traces supported PyTorch and NumPy computation across existing functions and objects. Export the complete path as one ONNX, TorchScript, or PT2 model, or choose separate stages when useful.

Input and output annotations capture computation as your application runs. A single model can capture computation across multiple input call sites, helper functions, and class methods. The annotations work directly in application code without a prescribed RL environment interface. See annotation patterns.

• Keep processing with the policy. Capture supported NumPy preprocessing, conversions into PyTorch, and action transforms together. Warp kernel regions can also be captured into ONNX or PT2 with an additional runtime adapter.
• Adapt unsupported calls with FunctionPatch. Register a faithful replacement built from supported operations; existing module-level callers can stay in place.
• Describe the deployment interface. Attach tensor roles, joint names, and timing information with semantic metadata, so downstream applications can interpret model inputs and outputs.
• Inspect and validate the export. Compare exported outputs, including state updates, against cached samples from consecutive Python iterations. Configure sample coverage, inspect the graph, and use sequential replay to check accumulated drift and runtime feedback.

Quick start

pip install leapp

Requires Python 3.10+ and PyTorch 2.7.0+.

This example exports NumPy preprocessing, a small PyTorch policy, and action scaling and clipping together:

import numpy as np
import torch
import leapp
from leapp import annotate

def make_observation(position, velocity):
    return np.concatenate([np.clip(position, -1.0, 1.0), velocity * 0.1])

policy = torch.nn.Sequential(
    torch.nn.Linear(12, 32), torch.nn.Tanh(), torch.nn.Linear(32, 6),
).eval()

leapp.start(name="robot_policy")
position, velocity = annotate.input_tensors("control", {
    "position": np.zeros(6, dtype=np.float32),
    "velocity": np.ones(6, dtype=np.float32),
})

with torch.no_grad():
    observation = torch.from_numpy(make_observation(position, velocity))
    targets = torch.clamp(policy(observation) * 0.25, -1.0, 1.0)

annotate.output_tensors("control", {"targets": targets}, export_with="onnx")

leapp.stop()
leapp.compile_graph(validate=True)

Produces robot_policy/control.onnx, a YAML interface specification, and a graph PNG on Python 3.11+. This stateless model runs directly in ONNX Runtime; InferenceManager loads the bundle for Python replay. Other applications can consume the exported artifacts through their chosen backend runtime.

Capture scope

LEAPP captures the executed tensor path. Data-dependent Python branches are not preserved dynamically, and unsupported operations or scalar extraction can freeze values. Check tracing warnings and validate on fresh inputs. See PyTorch and NumPy limitations.

Getting started · Export options · API reference

License

Copyright (c) 2026, NVIDIA Corporation & affiliates. Licensed under the Apache License, Version 2.0. See LICENSE and thirdparty.txt.