ONNX Runtime - Robotics AI

Important

ONNX Runtime support for Genio platforms is under active development. MediaTek currently provides more mature and comprehensive hardware-acceleration coverage for FP16 models. Support for QDQ (INT8) quantization is evolving, and performance will continue to improve as MediaTek expands operator coverage in quarterly updates.

This page provides models validated or planned on Genio platforms for robotic workloads, organized along the robot pipeline: perceive, interact, navigate, listen/speak, localize, and understand 3D. The resources support the Online inference path via ONNX Runtime.

Analytical AI inference paths: ONNX Runtime

ONNX Runtime is available on MediaTek Genio platforms to accelerate ONNX models with hardware support.

  • High Performance: Utilize the Neuron Execution Provider (EP) on Genio 520 and 720 for NPU acceleration.

  • Broad Compatibility: Standard CPU EP execution is available across all Genio platforms.

  • FP16 Focus: FP16 models have the most complete hardware-acceleration support.

  • Quantized QDQ: QDQ models have partially supported with NPU acceleration.

For best performance and compatibility, deploy FP16 ONNX models whenever possible.

Performance Comparison

MediaTek provides a comprehensive matrix of performance data across different Genio platforms.

  • To view a quick summary of high-end platforms (G520/G720), see the tables below.

  • To compare performance across all platforms (G350, G510, G700, G1200), refer to the dedicated page:

View Full Platform Performance Matrix

Robotics Capability Map

The 39 models are organized by robotic capability:

Category

What it enables

Representative models

Perception

Detect and classify objects in the robot’s camera view

YOLOv8, YOLO11n, RT-DETR, ViT

Interaction

Understand people: faces, body pose, text, and gaze

VGGFace, RTMPose, PaddleOCR

Navigation

Understand free space and geometry: segmentation, depth, and vision foundation models

Fast-SCNN, Depth Anything v2/v3, DINOv2

Audio

Speech-to-text, text-to-speech, and ambient sound processing

Moonshine tiny, ESPnet2, YAMNet

SLAM

Tracking, place recognition, and feature matching for localization and mapping

BoxMOT, MobileNetVLAD, SuperPoint

Point Cloud

3D object detection on LiDAR point clouds

PointPillars-Tiny, CenterPoint-Tiny

Others

Additional perception models that do not fit the categories above

UniGaze

Note

The performance statistics shown in these tables were measured using NPU Execution Provider with performance mode enabled via onnxruntime_perf_test.

The following symbols are used in the tables below:

  • -- : To be released.

  • X : Platform does not support this model.

Perception

Detect and classify objects in the robot’s camera view.

Models for Perception

Task

Model Name

Data Type

Input Size

G520 NPU (ms)

G520 CPU (ms)

G720 NPU (ms)

G720 CPU (ms)

Pretrained Model

Object detection

YOLOv5 [1]

–

–

–

–

–

–

–

Object detection

YOLOv8 [1]

–

–

–

–

–

–

–

Object detection

YOLO11n [1]

–

–

–

–

–

–

–

Object detection

MobileNetV2

–

–

–

–

–

–

–

Object detection

SqueezeNet

–

–

–

–

–

–

–

Object detection

ResNet50

–

–

–

–

–

–

–

Object detection

MobileNetV3

–

–

–

–

–

–

–

Object detection

RT-DETR v2/v4

–

–

–

–

–

–

–

Object detection

RF-DETR

–

–

–

–

–

–

–

Image classification

ViT

–

–

–

–

–

–

–

Interaction

Understand people: faces, body pose, text, and gaze.

Models for Interaction

Task

Model Name

Data Type

Input Size

G520 NPU (ms)

G520 CPU (ms)

G720 NPU (ms)

G720 CPU (ms)

Pretrained Model

Face recognition

VGGFace

–

–

–

–

–

–

–

Face recognition

MobileNetV3 embedding

–

–

–

–

–

–

–

Pose detection

MoveNet Lightning

–

–

–

–

–

–

–

Pose detection

RTMPose

–

–

–

–

–

–

–

Pose detection

HRNetPose

–

–

–

–

–

–

–

Pose detection

BlazePose Lite

–

–

–

–

–

–

–

Text recognition

PaddleOCR

–

–

–

–

–

–

–

Audio

Speech-to-text, text-to-speech, and ambient sound processing.

Models for Audio

Task

Model Name

Data Type

Input Size

G520 NPU (ms)

G520 CPU (ms)

G720 NPU (ms)

G720 CPU (ms)

Pretrained Model

STT

Moonshine tiny (streaming)

–

–

–

–

–

–

–

TTS

ESPnet2

–

–

–

–

–

–

–

Ambient sound: echo cancellation

TBD (Coming soon)

–

–

–

–

–

–

–

Ambient sound: noise reduction

TBD (Coming soon)

–

–

–

–

–

–

–

Ambient sound: sound detection

YAMNet

–

–

–

–

–

–

–

SLAM

Tracking, place recognition, and feature matching for localization and mapping.

Models for SLAM

Task

Model Name

Data Type

Input Size

G520 NPU (ms)

G520 CPU (ms)

G720 NPU (ms)

G720 CPU (ms)

Pretrained Model

Object tracking

BoxMOT

–

–

–

–

–

–

–

Localization

MobileNetVLAD

–

–

–

–

–

–

–

Localization

HFNet

–

–

–

–

–

–

–

Localization

SuperPoint

–

–

–

–

–

–

–

Localization

LightGlue

–

–

–

–

–

–

–

Point Cloud

3D object detection on LiDAR point clouds.

Models for Point Cloud

Task

Model Name

Data Type

Input Size

G520 NPU (ms)

G520 CPU (ms)

G720 NPU (ms)

G720 CPU (ms)

Pretrained Model

Point cloud detection

PointPillars-Tiny

–

–

–

–

–

–

–

Point cloud detection

CenterPoint-Tiny

–

–

–

–

–

–

–

Others

Additional perception models that do not fit the categories above.

Models for Others

Task

Model Name

Data Type

Input Size

G520 NPU (ms)

G520 CPU (ms)

G720 NPU (ms)

G720 CPU (ms)

Pretrained Model

Gaze estimation

UniGaze

–

–

–

–

–

–

–

Performance Notes and Limitations

Note

The measurements were obtained using onnxruntime_perf_test, and each model’s performance can vary depending on:

  • The specific Genio platform and hardware configuration.

  • The version of the board image and evaluation kit (EVK).

  • The selected backend and model variant.

To obtain the most accurate performance numbers for your use case, you must run the application directly on the target platform.

Current Limitations

ONNX-GAI (generative AI) models are not officially supported at this time. Some proof-of-concept experiments exist internally, but they are not production-ready and are not part of the validated model list.

For the latest roadmap or early-access updates on ONNX-GAI support, please refer to the Genio Community forum.