LiteEMG-FM: An Efficient and Deployable Foundation Model for Robust EMG Sensing
Authors: Tianhao Wu, Xu Wu, Amirmohammad Radmehr, Jiawei Yu, Yi Wu, Phuc Nguyen, Jian Liu
Organizations: University of Georgia, Athens, Georgia, United States · Computer Science, University of Oklahoma, Norman, Oklahoma, United States · Wireless and Sensor Systems Lab (WSSL), University of Massachusetts Amherst, Amherst, Massachusetts, United States · Manning College of Information and Computer Sciences, University of Massachusetts Amherst, Amherst, Massachusetts, United States
Electromyography (EMG) signals vary substantially across individuals, body regions, recording sessions, and sensing hardware, limiting the generalization of models for assistive devices and human-computer interaction. Existing time-series foundation models are also computationally expensive for real-time wearable deployment and often fail to capture EMG-specific time-frequency characteristics. We present LiteEMG-FM, an efficient hybrid CNN-Transformer foundation model for practical EMG sensing. Pretrained on 16 diverse upper- and lower-limb EMG datasets, LiteEMG-FM learns representations that generalize across users and datasets. For resource-constrained deployment, we implement a hierarchical wake-up architecture in which a lightweight, always-on 1D-CNN filters rest and non-target activity and activates LiteEMG-FM only for valid gestures. We evaluate full inference offloading, split inference, and full on-device processing, characterizing their trade-offs in latency, power consumption, and memory footprint. Across diverse evaluation settings, LiteEMG-FM outperforms state-of-the-art time-series foundation models and supervised baselines, particularly under zero-calibration cross-participant and data-scarce conditions. These results demonstrate that LiteEMG-FM is an effective, efficient, and deployable foundation model for EMG applications.
Figures & tables
Figure 1. LiteEMG-FM : the first EMG foundation model explicitly designed to balance strong generalization with wearable deployment constraints.
Figure 2. Overview of LiteEMG-FM . Raw EMG signals are processed into time–frequency spectrograms, encoded by a hybrid CNN–Transformer model, pretrained via masked reconstruction, and fine-tuned for hand and leg gesture classification.
Dataset
Body Part
SP (Hz)
Channels
Task
Hand/Arm Gesture Datasets
NinaPro DB1 ( Atzori et al., 2014 )
Forearm
100
10
Hand Gesture Classification
NinaPro DB2–4 / DB6–8 ( Atzori et al., 2014 ; Pizzolato et al., 2017 ; Krasoulis et al., 2017 ; Palermo et al., 2017 ; Krasoulis et al., 2017 ; Krasoulis et al., 2019 )
Hand
2000
10/12
Hand Gesture Classification
NinaPro DB5 ( Pizzolato et al., 2017 )
Forearm
200
10
Hand Gesture Classification
GRABMyo ( Jiang et al., 2022 )
Forearm
2048
32
Hand Gesture Classification
CapgMyo ( Dai et al., 2021a )
Hand
1000
128
Hand Gesture Classification
Table 1. Summary of EMG datasets used in LiteEMG-FM .
Figure 3. Overview of representative activities in the pretraining dataset. The left panel shows representative hand gestures (e.g., grasping, pointing, finger counting). The right panel shows locomotion activities (e.g., walking, marching, stair climbing). The datasets span 3 to 52 gesture classes, providing diverse muscle activation patterns for pretraining and downstream tasks.
Model
Year
Type
Backbone
Primary Tasks
# of Parameters
LiteEMG-FM (Ours)
2026
Foundation (Self-Supervised)
CNN + Transformer Encoder
Efficient EMG encoding for gesture, locomotion
8 M
EMGNet ( Chen et al., 2020 )
2020
Supervised
2D CNN
Hand gesture classification
2 M
LSTM-RNN ( Toro-Ossaba et al., 2022 )
2022
Supervised
LSTM-RNN
Hand gesture classification
1 M
WaveFormer ( Chen et al., 2025a )
2025
Supervised
Transformer
Hand gesture classification
7 M
MOMENT-Small ( Goswami et al., 2024 )
2024
Foundation (Self-Supervised)
Transformer
General time-series representation learning
40 M
MOMENT-Base ( Goswami et al., 2024 )
2024
Foundation (Self-Supervised)
Transformer
General time-series representation learning
125 M
Table 2. Comparison of baseline models and the proposed LiteEMG-FM .
Dataset
LiteEMG-FM
MOMENT-S
MOMENT-B
MOMENT-L
Brant
PatchTST
EMGNet
WaveFormer
LSTM-RNN
NinaPro DB2 (hand)
69.2 / 69.1
45.2 / 44.2
48.1 / 47.6
62.1 / 61.6
35.9 / 34.7
41.7 / 40.3
44.2 / 41.9
56.5 / 56.6
44.6 / 43.5
NinaPro DB7 (hand)
79.8 / 76.3
53.7 / 48.2
55.0 / 50.7
65.9 / 61.9
36.1 / 32.1
48.7 / 44.0
48.7 / 40.5
65.7 / 63.4
34.8 / 31.6
NinaPro DB8 (hand)
81.7 / 82.5
45.6 / 45.0
44.1 / 44.8
59.7 / 60.4
37.2 / 34.6
48.9 / 48.9
64.1 / 63.2
74.2 / 74.8
44.6 / 43.9
SIAT-LLMD (leg)
53.6 / 59.0
41.9 / 48.7
41.1 / 46.2
40.0 / 45.4
29.5 / 28.3
40.1 / 46.2
49.0 / 55.2
48.4 / 53.1
33.2 / 33.9
MyPredict1 (leg)
97.2 / 94.9
68.4 / 30.8
85.8 / 70.7
93.7 / 85.4
74.3 / 78.5
85.1 / 76.0
61.5 / 39.3
35.6 / 16.9
87.8 / 73.4
Table 3. Downstream performance across five datasets. LiteEMG-FM and other foundation models are trained with full fine-tuning; supervised models are trained from scratch. Values are percentages (Acc./F1).
Dataset
LiteEMG-FM
MOMENT-S
MOMENT-B
MOMENT-L
Brant
PatchTST
NinaPro DB2 (hand)
64.6 / 64.2
12.8 / 10.5
18.5 / 18.0
23.0 / 19.5
19.9 / 18.9
26.5 / 22.8
NinaPro DB7 (hand)
66.1 / 62.0
17.0 / 9.2
20.9 / 11.6
23.5 / 12.7
18.9 / 13.4
31.3 / 19.8
NinaPro DB8 (hand)
80.5 / 81.3
20.4 / 11.1
22.4 / 14.9
23.8 / 13.7
29.9 / 23.4
22.1 / 20.9
SIAT-LLMD (leg)
48.8 / 53.9
22.8 / 21.5
25.6 / 29.3
24.9 / 25.0
21.7 / 22.2
21.1 / 15.7
MyPredict1 (leg)
91.2 / 83.2
69.0 / 31.3
74.4 / 36.9
73.5 / 31.4
73.1 / 45.2
55.6 / 10.3
Table 4. Comparison between LiteEMG-FM (Linear Probe) and baseline foundation models across five downstream datasets. All values are reported in percentages (Acc./F1).
Dataset
Model
5%
10%
20%
40%
80%
100%
NinaPro DB2 (hand)
LiteEMG-FM
23.0
29.4
37.6
48.0
61.7
64.5
MOMENT-S
14.5
21.3
35.4
42.7
53.6
56.5
MOMENT-B
14.3
22.5
33.0
43.9
55.1
59.2
MOMENT-L
15.3
26.5
34.7
41.7
57.5
62.1
Brant
9.9
12.9
16.2
23.1
30.4
33.0
PatchTST
7.4
12.9
15.8
28.0
38.8
41.5
Table 5. Accuracy under label scarcity for LiteEMG-FM and baseline models on MyPredict (leg) and NinaPro DB2 (hand). Values are shown in percentages. LiteEMG-FM uses linear probing; baseline CNNs and Transformers are trained from scratch.
Model
Acc/F1
Model
Acc/F1
Model
Acc/F1
LiteEMG-FM
40.0/39.0
MOMENT-L
39.2/37.7
EMGNet
24.6/23.1
MOMENT-S
34.1/33.8
Brant
18.7/18.4
WaveFormer
17.2/13.5
MOMENT-B
36.2/35.5
PatchTST
32.5/32.0
LSTM-RNN
17.2/15.8
Table 6. Comparison of cross-subject classification accuracy.
Model Configuration
NinaPro DB2
NinaPro DB7
NinaPro DB8
SIAT-LLMD
MyPredict1
LiteEMG-FM (Full Model)
69.2
79.8
81.7
53.6
97.2
- Mask Ratio 0.3 (vs. 0.5)
64.5
74.6
81.1
52.6
95.4
- 4 Encoder Layers (vs. 2)
72.6
81.7
82.4
54.7
96.7
Table 7. Ablation study on pretraining design. Performance is Linear Probing Accuracy across five downstream datasets.
Figure 4. Experimental hardware implementation. The edge node consists of an ESP32-S3 microcontroller interfacing with raw EMG sensors. For Scenarios A and B, the device transmits data (raw signals or latent embeddings) via Bluetooth Low Energy to a companion Server for offloaded inference. This setup is used to empirically measure the latency, memory, and power for each deployment strategy.
Figure 5. Overview of the LiteEMG-FM hierarchical deployment architecture. To minimize energy waste, an always-on TinyCNN Gate runs locally on the edge device to filter rest intervals and motion noises. Upon detecting valid gestures, the system executes one of three inference paradigms: Scenario A Full Inference Offloading streams raw data to a server to minimize edge computing; Scenario B, Split Inference executes the convolutional tokenizer locally and transmits compressed latent embeddings to balance bandwidth and compute, and Scenario C Full On-Device Processing executes the entire foundation model locally to maximize data privacy and eliminate network dependence.
Specification
Value
Microcontroller
Espressif ESP32-S3-WROOM-1
CPU Core
Dual-core Xtensa® LX7 @ 240 MHz
Instruction Set
32-bit RISC with Vector Extensions
Memory (RAM)
512 KB Internal SRAM + 8 MB Octal PSRAM
Storage
8 MB SPI Flash
Operating System
FreeRTOS (via ESP-IDF v5.x)
Table 8. Hardware specifications of the wearable edge platform.
Deployment Strategy
Latency (ms)
Memory Footprint
Energy
Edge
Server
Total
Flash
Peak RAM
Active Power
A: Full Offloading
60
5
65
< 100 KB
≈ 240 KB
240 mW
B: Split Inference
5,928
1.7
5,930
1.5 MB
1,700 KB
387 mW
C: Full Edge
16,912
0.0
16,912
7.8 MB
2,000 KB
422 mW
Table 9. System performance trade-offs across deployment scenarios. Metrics are measured per 3-second inference window.
EMG-based gesture recognition decodes hand and arm movements from muscle electrical signals for prosthetic control, wearables, and human-machine interfaces.
Imperial College London London, United Kingdom · Nokia Bell Labs London, United Kingdom · Brave Software & University of Cambridge London, United Kingdom