ACT: Optimized for Qualcomm Devices

ACT (Action Chunking with Transformers) is a robotic policy model that is trained to predict the next chunk of actions that the robotic hand is expected to perform.

This is based on the implementation of ACT found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.42, ONNX Runtime 1.24.3	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.19.1	Download

For more device-specific assets and performance metrics, visit ACT on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

• Custom weights (e.g., fine-tuned checkpoints)
• Custom input shapes
• Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for ACT on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.robotics

Model Stats:

• Model checkpoint: act
• Input resolution: 480x640
• Number of parameters: 83.92M
• Model size (float): 255M

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
ACT	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	5.518 ms	4 - 334 MB	NPU
ACT	ONNX	float	Snapdragon® X2 Elite	6.176 ms	63 - 63 MB	NPU
ACT	ONNX	float	Snapdragon® X Elite	11.87 ms	62 - 62 MB	NPU
ACT	ONNX	float	Snapdragon® 8 Gen 3 Mobile	8.053 ms	2 - 377 MB	NPU
ACT	ONNX	float	Qualcomm® QCS8550 (Proxy)	11.249 ms	0 - 81 MB	NPU
ACT	ONNX	float	Qualcomm® QCS9075	18.602 ms	4 - 10 MB	NPU
ACT	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	6.567 ms	1 - 331 MB	NPU
ACT	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	3.739 ms	3 - 321 MB	NPU
ACT	QNN_DLC	float	Snapdragon® X2 Elite	4.903 ms	4 - 4 MB	NPU
ACT	QNN_DLC	float	Snapdragon® X Elite	8.9 ms	4 - 4 MB	NPU
ACT	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	5.957 ms	0 - 345 MB	NPU
ACT	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	43.945 ms	2 - 309 MB	NPU
ACT	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	8.287 ms	4 - 6 MB	NPU
ACT	QNN_DLC	float	Qualcomm® SA8775P	13.344 ms	1 - 294 MB	NPU
ACT	QNN_DLC	float	Qualcomm® QCS9075	15.692 ms	4 - 9 MB	NPU
ACT	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	16.31 ms	0 - 260 MB	NPU
ACT	QNN_DLC	float	Qualcomm® SA7255P	43.945 ms	2 - 309 MB	NPU
ACT	QNN_DLC	float	Qualcomm® SA8295P	15.499 ms	0 - 231 MB	NPU
ACT	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	4.795 ms	0 - 297 MB	NPU
ACT	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	3.799 ms	0 - 327 MB	NPU
ACT	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	5.992 ms	0 - 360 MB	NPU
ACT	TFLITE	float	Qualcomm® QCS8275 (Proxy)	44.193 ms	0 - 319 MB	NPU
ACT	TFLITE	float	Qualcomm® QCS8550 (Proxy)	8.252 ms	0 - 2 MB	NPU
ACT	TFLITE	float	Qualcomm® SA8775P	13.455 ms	0 - 304 MB	NPU
ACT	TFLITE	float	Qualcomm® QCS9075	16.342 ms	0 - 71 MB	NPU
ACT	TFLITE	float	Qualcomm® QCS8450 (Proxy)	16.296 ms	0 - 275 MB	NPU
ACT	TFLITE	float	Qualcomm® SA7255P	44.193 ms	0 - 319 MB	NPU
ACT	TFLITE	float	Qualcomm® SA8295P	15.755 ms	0 - 235 MB	NPU
ACT	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	4.845 ms	0 - 309 MB	NPU

License

• The license for the original implementation of ACT can be found here.

References

• Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.