autoware_simpl_prediction

Purpose

The autoware_simpl_prediction is used for 3D object motion prediction based on ML-based model called SIMPL.

Inner-workings / Algorithms

The implementation bases on SIMPL [1] [2] work. It uses TensorRT library for data process and network interface.

Workflow overview of this node is as follows:

flowchart TD
    In1@{ shape: card, label: "~/input/objects" } -- autoware_perception_msgs::msg::TrackedObjects --> Callback1@{ shape: rect, label: "SimplNode::callback(...)" }
    Callback1 --> B@{ shape: stadium, label: "Start measuring Processing Time" }

    In2@{ shape: card, label: "/localization/kinematic_state" } -- nav_msgs::msg::Odometry --> C
    B --> C@{ shape: subproc, label: "SimplNode::subscribe_ego(...)" }

    In3@{ shape: card, label: "~/input/vector_map" } -- "autoware_map_msgs::msg::LaneletMapBin" --> Callback2@{ shape: rect, label: "SimplNode::on_map(...)" }
    Callback2 --> X@{ shape: subproc, label: "LaneletConverter::convert(...)" }

    C -->|✅| D@{ shape: subproc, label: "LaneletConverter::polylines()" }
    C -->|❌| Z1@{ shape: curv-trap, label: "⚠️WARNING: Failed to subscribe ego" } --> END@{ shape: stadium }

    D -->|✅| E@{ shape: subproc, label: "SimplNode::update_history(...)" }
    D -->|❌| Z2@{ shape: curv-trap, label: "⚠️WARNING: No map points" } --> END

    E --> F@{ shape: subproc, label: "PreProcessor::process(...)" }
    F --> G@{ shape: subproc, label: "TrtSimpl::do_inference(...)" }

    G -->|✅| I@{ shape: subproc, label: "PostProcessor::process(...)" } --> J@{ shape: stadium, label: "Publish Predicted Objects" }
    G -->|❌| Z3@{ shape: curv-trap, label: "⚠️ERROR: Inference failed" } --> H

    J -- autoware_perception_msgs::msg::PredictedObjects --> Out1@{ shape: card, label: "~/output/objects" }
    J --> H@{ shape: stadium, label: "Publish Processing & Cyclic Time" }

    H -- autoware_internal_debug_msgs::msg::Float64Stamped --> Out3@{ shape: card, label: "~/debug/cyclic_time_ms" }
    H -- autoware_internal_debug_msgs::msg::Float64Stamped --> Out2@{ shape: card, label: "~/debug/processing_time_ms" }

    H --> END

Inputs Representation

• $X_A\in R^{N\times D_{agent}\times T_{past}}$: Agent histories input.
• $X_M\in R^{K\times P\times D_{map}}$: Map points input.
• $X_{RPE}\in R^{(N+K)\times (N+K)\times D_{rpe}}$: Relative pose encoding input.

Outputs Representation

• $P_{score}\in R^{N\times M}$: Predicted scores for each agent and mode.
• $P_{trajectory}\in R^{N\times M\times T_{future}\times D_{trajectory}}$: Predicted trajectories for each agent and mode.
  • Where, $D_{trajectory}$ is $(x, y, v_x, v_y)$ in the agent local coordinate frame.

Inputs / Outputs

Inputs

Outputs

Debug Outputs

Parameters

{{ json_to_markdown(“perception/autoware_simpl/schema/simpl.scheme.json”) }}

[WIP] Model Training / Deployment

Now we are preparing a library to train and deploy SIMPL and other ML models featuring motion prediction tasks.

Testing

Unit tests are provided and can be run with:

colcon test --packages-select autoware_simpl
colcon test-result --all

To print the test’s details with while the tests are being run, use the --event-handlers console_cohesion+ option to print the details directly to the console:

```bash colcon test –event-handlers console_cohesion+ –packages-select autoware_simpl

File truncated at 100 lines see the full file

Changelog for package autoware_simpl_prediction

0.47.0 (2025-08-11)

• feat(autoware_simpl): add ML-based motion prediction node (#10824)
• Contributors: Kotaro Uetake