Open Source RoboticsRobots, software & learning

Intrinsic Inference

Software

Preview framework for managing model-serving lifecycles and exposing inference services to ROS 2. [1]

Installation & first result →Official websiteGetting started

For
ROS 2 developers integrating an inference backend [1]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
operating system, compute requirements not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
Apache-2.0 · license not fully assessed for model weights · Details by part

Requirements · Research details and dates

On this page

Overview

Intrinsic Inference manages model staging, loaded-model state and request forwarding to an inference backend. Its ROS node exposes that machinery through services. It does not supply a trained robotics policy just by installing the framework, and the repository explicitly labels the current work a preview that is not stable. [1] [3] [4]

Is this for you?

Workflows
  • Stage and reconcile model assets
  • Proxy inference through a backend
  • Query backend readiness and model metadata from ROS

[3] [4]

Known limitations
  • Upstream explicitly labels this a preview that is not stable

[1]

What you need

Dependencies
  • Bazel
  • NVIDIA Triton server; quickstart uses Docker image 26.07-py3
  • CycloneDDS for the documented ROS client/server route
  • A separately provided model repository

[3]

Installation & first result

Open the upstream getting-started guide [1] · Project documentation [1]

  • Create the documented model directory and fetch its external config/weights
  • Start Triton with --model-control-mode=explicit
  • Launch inference_node_main with repo_path and triton_grpc_url
  • Build/source the custom ROS interfaces and align the client middleware
  • Call server_live and inspect model metadata before attempting inference

[3]

Installation

Prepare the example model repository, start Triton in explicit model-control mode, then launch the Bazel InferenceNode target and make its custom interfaces visible to the ROS client.

[3]

Hardware, models & data

Inputs and outputs

Model repository manifests/files and inference requests feed a lifecycle manager and gRPC backend proxy; ROS services expose liveness, readiness, metadata and inference results.

[3] [4]

File formats
  • ONNX in the demo
  • Protocol Buffers model manifests
  • ROS service definitions

[3] [4]

Project notes

Start the backend before the ROS wrapper

The ROS quickstart prepares an external DenseNet ONNX model directory and starts Triton using the 26.07-py3 container image. Triton must run with --model-control-mode=explicit, because the framework manages model loading and unloading itself. It then launches inference_node_main with the model repository path and Triton's gRPC address. [3]

For an initial check, make the custom inference interfaces available in the client workspace, source that workspace and call the documented server_live service. Querying model metadata is the next check before attempting useful inference. The guide selects CycloneDDS for both the Bazel-built node and its external ROS client; do not assume the Zenoh settings from Core's visualization example apply here. [3]

Serving code is separate from the selected model

The core reconciles desired and actual model state, stages files and proxies requests. Compute demand therefore depends on the chosen model and backend; this review establishes no universal GPU, memory or throughput figure. The software repository carries Apache-2.0, but that does not resolve the rights of the separately downloaded model. [2] [3] [4]

This profile pins an unreleased commit snapshot. We have not started Triton, downloaded the example weights or measured inference through the ROS wrapper. [90]

What we checked

Source research only. No installation, upstream test execution, physical integration or performance measurement by this publication. No affiliation or endorsement claimed.

Not yet established

  • Requirements: operating system, host computer
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Sep 23, 2026 [90]
Latest release
Not established
Profile updated
Oct 4, 2026
Sources last checked
Sep 30, 2026
Next review
Oct 30, 2026
Reviewed revision
7760abb8097484a8168e34848403ca4b165e4480 (commit snapshot; no published stable release found) · software repository

Source files and licenses

Useful open software component with documented development workflow; dependency rights, configuration matching and runtime results remain separate.

LayerAvailability / license
softwareavailable · Apache-2.0
Inspected root license for this repository. External dependencies, vendor SDKs, models and third-party geometry are not relicensed by it. [1] [2]
model weightsavailable · License not fully assessed
The quickstart downloads a separate DenseNet ONNX model. Exact model licensing and workload qualification are not established by this software review. [3]
documentationavailable · Apache-2.0
Inspected root license for this repository. External dependencies, vendor SDKs, models and third-party geometry are not relicensed by it. [1] [2]

No particular model weights or dataset is offered or assessed by this component profile.

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Project sources

Videos

  • S091 Intrinsic (YouTube channel)
    The project’s own YouTube channel, found by searching YouTube and checked against the channel page. Not reviewed by us for accuracy.
    Revision, date & where cited

    YouTube channel Intrinsic · youtube.com · Accessed Oct 4, 2026

Correction guidance · Discussion is not enabled · Profile format 0.6.0-draft.1

Change history

2026-09-30 — Initial pinned-source component profile with setup route and integration boundaries.

2026-10-04 — Added 1 YouTube reference(s) (S091): official channel or videos linked from the pinned upstream sources. Existing claims, source-check dates and evidence are unchanged.