SIMPL for Autoware (simpl_prediction)

Multi-agent motion prediction model for autonomous driving, used by the autoware_simpl_prediction node in Autoware.

The model follows the SIMPL [1] architecture (A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving) and runs with TensorRT inside Autoware. It is exported as ONNX so it can be deployed across hardware; Autoware builds the TensorRT engine from the ONNX file on first launch.

Model overview

Task	Multi-modal trajectory prediction for tracked traffic agents
Architecture	SIMPL [1]
Predicted agent classes	VEHICLE, PEDESTRIAN, MOTORCYCLIST, CYCLIST, LARGE_VEHICLE
Runtime	TensorRT (FP32 by default) via the autoware_simpl_prediction ROS 2 node
Format	ONNX (Autoware builds the TensorRT engine locally on first launch)
License	Apache-2.0

Autoware object labels are mapped to the model's agent classes as follows: CAR to VEHICLE, PEDESTRIAN to PEDESTRIAN, BICYCLE to CYCLIST, MOTORCYCLE to MOTORCYCLIST, and TRUCK / TRAILER / BUS to LARGE_VEHICLE. Only agents whose mapped label is listed in the node's preprocess.labels parameter are predicted and published.

The following dimensions are fixed at ONNX export time (dynamic shape inference is not supported yet), and the default node configuration matches them:

Parameter	Symbol	Value
Maximum number of agents	N	50
Past history length	T_past	8
Maximum number of map polylines	K	300
Maximum points per polyline	P	10
Number of prediction modes	M	6
Future horizon length	T_future	80

Files

File	Description
simpl.onnx	SIMPL motion prediction network (ONNX)
deploy_metadata.yaml	Deployment metadata recording the artifact version of this repository

TensorRT engines are not distributed here. TensorRT engines are specific to the GPU architecture and TensorRT version they are built on and are not portable, so Autoware builds them locally from the ONNX file on first launch (or via build_only:=true).

Inputs and outputs

Network inputs

• Agent histories: N x D_agent x T_past (past states of the tracked agents)
• Map points: K x P x D_map (lanelet map polylines)
• Relative pose encoding: (N + K) x (N + K) x D_rpe

Network outputs

• Predicted scores: N x M (confidence per agent and mode)
• Predicted trajectories: N x M x T_future x D_trajectory, where D_trajectory is (x, y, vx, vy) in the agent local coordinate frame

Node topics (as used by autoware_simpl_prediction)

Topic	Type	Direction
~/input/objects	autoware_perception_msgs/msg/TrackedObjects	input
~/input/vector_map	autoware_map_msgs/msg/LaneletMapBin	input
/localization/kinematic_state	nav_msgs/msg/Odometry	input
~/output/objects	autoware_perception_msgs/msg/PredictedObjects	output

Pre-processing (agent history accumulation, lanelet-to-polyline conversion) and post-processing (mode score thresholding) run in the node, not in the ONNX graph.

Usage in Autoware

The node expects the artifacts under ~/autoware_data/ml_models/simpl_prediction/ and launches with:

ros2 launch autoware_simpl_prediction simpl.launch.xml

Add build_only:=true to build the TensorRT engine from the ONNX as a one-off pre-task. The default parameters (config/simpl.param.yaml) point to simpl.onnx in the data directory and use fp32 precision. See the package README for the full parameter reference.

Training

The SIMPL architecture and reference training code originate from the upstream project:

• Original implementation: https://github.com/HKUST-Aerial-Robotics/SIMPL
• Paper: "SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving", arXiv:2402.02519

An Autoware library to train and deploy SIMPL and other motion prediction models is work in progress, as noted in the package README. The training data and training configuration used to produce the released simpl.onnx weights are not publicly documented.

Limitations

• The number of predictable agents is fixed at export time (max_num_agent: 50); dynamic shape inference is not supported yet.
• Only agents whose label is contained in preprocess.labels are published; other road users are ignored.
• Predicted modes with a confidence score below postprocess.score_threshold are filtered out. If all modes of an object are filtered, the published object contains no path, and the remaining mode confidences are not guaranteed to sum to 100%.
• Agent history that is no longer observed in incoming callbacks is dropped, which resets prediction context for reappearing agents.

Provenance

Original artifact source	https://awf.ml.dev.web.auto/perception/models/simpl/v0.1.0/simpl.onnx
Source version string	simpl/v0.1.0
Tag in this repository	v0.1 (normalized from v0.1.0)
simpl.onnx SHA-256	ad5e03983193c4d188432f314334697d6a216e7ffc91fb651eee5d6e4c42f492

Citation

@article{zhang2024simpl,
  title   = {SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving},
  author  = {Zhang, Lu and Li, Peiliang and Liu, Sikang and Shen, Shaojie},
  journal = {arXiv preprint arXiv:2402.02519},
  year    = {2024}
}

References

• [1] Zhang et al., "SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving", arXiv:2402.02519, 2024.
• [2] Original implementation: https://github.com/HKUST-Aerial-Robotics/SIMPL