EmbodiTTA: Resource-Efficient Test-Time Adaptation for Embodied Visual Systems

Xiao Ma; Young D. Kwon; Dong Ma

cs.LGMay 2, 2025

EmbodiTTA: Resource-Efficient Test-Time Adaptation for Embodied Visual Systems

Authors: Xiao Ma, Young D. Kwon, Dong Ma

Organizations: Singapore Management University, Singapore · Samsung AI Center–Cambridge, United Kingdom · University of Cambridge, United Kingdom

Abstract

Continual Test-time adaptation (CTTA) continuously adapts the deployed model on every incoming batch of data. While achieving optimal accuracy, existing CTTA approaches present poor real-world applicability on resource-constrained edge devices, due to the substantial memory overhead and energy consumption. In this work, we first introduce a novel paradigm -- on-demand TTA -- which triggers adaptation only when a significant domain shift is detected. Then, we present OD-TTA, an on-demand TTA framework for accurate and efficient adaptation on edge devices. OD-TTA comprises three innovative techniques: 1) a lightweight domain shift detection mechanism to activate TTA only when it is needed, drastically reducing the overall computation overhead, 2) a source domain selection module that chooses an appropriate source model for adaptation, ensuring high and robust accuracy, 3) a decoupled Batch Normalization (BN) update scheme to enable memory-efficient adaptation with small batch sizes. Extensive experiments show that OD-TTA achieves comparable and even better performance while reducing the energy and computation overhead remarkably, making TTA a practical reality.

Figures & tables

Figure 1: Impact of domain shift: indoor $\rightarrow$ outdoor.

Panel of Figure 2 — Figure 2: Illustration of (a) continual and (b) on-demand test-time adaptation.

Panel of Figure 3 — Figure 3: (a) Correlation of accuracy and entropy, (b) adaptation to a target domain from different source domains.

Panel of Figure 4 — Figure 4: (a) TTA performance under different batch sizes, (b) memory usage of updating BN statistics and parameters.

Panel of Figure 6 — Figure 6: Sample-wise domain shift detection using (a) entropy and (b) the proposed EMA entropy.

Panel of Figure 7 — Figure 7: (a) Candidate domain construction, (b) storage comparison of saving only BN layers and the full model.

	Batch Size = 1			Batch Size = 16			Batch Size = 64
	Avg. Acc.(%)			Avg. Acc.(%)			Avg. Acc.(%)
Method	CIFAR10-C	ImageNet-C	Memory (MB)	CIFAR10-C	ImageNet-C	Memory (MB)	CIFAR10-C	ImageNet-C	Memory (MB)
Source	59.5	26.9	242	59.5	26.9	358	59.5	26.9	809
SAR	68.3	35.1	429	70.4	35.2	1723	69.2	37.3	5780
MECTA	65.0	12.0	378	81.3	33.2	1231	81.3	33.7	3836
ROID	10.1	0.4	698	79.5	32.5	2850	82.8	37.8	9840

Table I: Comparison of accuracy (%) on CIFAR10-C and ImageNet-C using ResNet-50. Memory is measured on Jetson Orin Nano whenever the method fits on the device; otherwise, it is measured on a server. Batch size 64 is measured on a server for all methods due to Jetson memory limitations.

Method	Batch size = 1	Batch size = 16	Batch size = 64
Source	69.9	69.9	69.9
SAR	76.8	75.9	74.6
MECTA	62.7	83.4	84.2
ROID	6.2	81.3	84.3
L-TTA	2.2	83.0	83.3
SURGEON	2.9	79.5	80.4

Table II: Comparison of accuracy (%) on CORe50 dataset using ResNet-50.

	SHIFT types
Method	Day $\rightarrow$ Night	Clear $\rightarrow$ Foggy	Clear $\rightarrow$ Rainy	Avg.
Source	27.3	17.7	11.0	18.7
SAR	23.6	7.6	4.2	11.8
MECTA	24.2	20.6	15.4	20.1
SURGEON	30.1	23.6	18.2	24.0
EATA-C	24.7	21.5	16.7	20.9

Table III: Comparison of Adaptation mIoU (%) on SHIFT using DeepLabV3.

Lengths	1000	3000	5000	7000	9000	10000
SAR	138 (1.2 $\times$ )	414	690	965	1241	1379 (1.2 $\times$ )
MECTA	160 (1.0 $\times$ )	480	800	1120	1440	1600 (1.0 $\times$ )
ROID	75 (2.1 $\times$ )	226	376	526	676	749 (2.1 $\times$ )
L-TTA	50 (3.2 $\times$ )	143	240	333	424	477 (3.4 $\times$ )
SURGEON	78 (2.1 $\times$ )	227	376	531	678	770 (2.1 $\times$ )
EATA-C	47 (3.4 $\times$ )	136	227	317	408	453 (3.5 $\times$ )

Table IV: Latency for processing domain data sequences of varying lengths under batch size of 1 on CIFAR10-C.

Panel of Figure 8 — Figure 8: Energy consumption for processing domain data sequences of varying lengths under batch size = (a) 1 and (b) 16.

Figure 9: EMA entropy change along the data stream on CIFAR10-C. Domain transitions occur every 10,000 samples and are highlighted in alternating background (white-gray) color. Red dotted lines indicate detected shifts. The table above shows the corresponding accuracy on each domain.

Sequences	S1	S2	S3	S4	S5	S6	S7	S8	S9	S10	Mean $\pm$ Std.
Triggered	15	13	15	14	13	14	15	13	13	14	13.9 $\pm$ 0.8
Missed	1	2	0	1	2	1	0	2	2	1	1.2 $\pm$ 0.7
False	1	0	0	0	0	0	0	0	0	0	0.1 $\pm$ 0.3
Delay	301	569	262	837	322	279	477	531	488	523	459 $\pm$ 167
Avg. acc.	79.9	79.3	80.5	79.5	79.5	80.2	80.2	78.7	77.7	79.7	79.5 $\pm$ 0.8

Table V: Impact of different domain shift orders (S1–S10). Each sequence is a random shuffle of the 15 domains.

Table 15

Adapt from.	Batch Size = 1		Batch Size = 16		Batch Size = 64
Adapt from.	CIFAR10-C	ImageNet-C	CIFAR10-C	ImageNet-C	CIFAR10-C	ImageNet-C
Prev-Domain	79.5	27.8	84.4	31.4	86.0	33.4
Initial-Pool	81.0	34.6	85.7	36.7	86.1	39.1
Progressive-Pool	81.7	34.5	85.1	37.5	86.0	37.7

Table VIII: Evaluation of source model selection on CIFAR10-C and ImageNet-C.

Method	MobileNetV2			MobileViT
Method	BS = 1	BS = 16	BS = 64	BS = 1	BS = 16	BS = 64
SAR	0.1	21.4	25.7	0.1	30.6	32.9
MECTA	0.2	0.6	0.7	0.2	0.2	0.2
ROID	0.2	21.5	25.4	0.2	30.0	33.1
SURGEON	0.2	1.38	23.5	0.2	2.51	28.4
EATA-C	0.1	21.8	26.2	0.9	27.3	32.5

Table IX: Adaptation accuracy (%) on ImageNet-C using MobileNetV2 and MobileViT.

Test-Time Adaptation

Test-time adaptation updates or recalibrates pretrained models at inference to handle distribution shifts, often without target labels.

See all 322 papers

Efficient Test-Time Adaptation through Latent Subspace Coefficients Search
Oct 13, 2025Xinyu Luo, Jie Liu, Kecheng Chen +6Test-Time Adaptation Continual Test-Time Adaptation
Department of Electrical Engineering, City University of Hong Kong, Hong Kong SAR, China
ZOTTA: Test-Time Adaptation with Gradient-Free Zeroth-Order Optimization
Mar 15, 2026Ronghao Zhang, Shuaicheng Niu, Qi Deng +3Test-Time Adaptation Zeroth-Order Optimization
South China University of Technology, Guangzhou, 510000, China. · Nanyang Technological University, 639798, Singapore. · Xi’an Jiaotong University, Xi’an, 710049, China. +1
Cross-device Collaborative Test-time Adaptation with Zeroth-order Optimization and Model Merging
Jul 3, 2026Yu Mitsuzumi, Akisato Kimura, Yasuhiro Fujiwara +1Test-Time Adaptation On-Device Inference
NTT, Inc. · Kyoto University

Oct 13, 2025J/K move · Enter open · S save

Efficient Test-Time Adaptation through Latent Subspace Coefficients Search

Xinyu Luo, Jie Liu, Kecheng Chen, Yihua Shao+5

Department of Electrical Engineering, City University of Hong Kong, Hong Kong SAR, China

Test-Time Adaptation Continual Test-Time Adaptation

Open paper AskPDF arXiv

Real-world deployment often exposes vision models to distribution shifts, making test-time adaptation (TTA) important for reliable prediction. Existing TTA methods can be difficult to deploy on edge devices because they require backpropagation, activation storage, or test-time batches, which increase computation, memory, and state. We propose ELaTTA (Efficient Latent Test-Time Adaptation), a forward-only, gradient-free framework for single-instance TTA. ELaTTA freezes the pretrained model and adapts each test input by searching a low-dimensional coefficient vector in a principal latent subspace computed offline from a source-domain cache. The basis is stored once and source data are discarded before deployment. At test time, CMA-ES performs covariance-adaptive Gaussian search, ranks candidate corrections by prediction entropy, and selects the best candidate within a fixed evaluation budget. On five visual benchmarks spanning corruption, natural, rendition, sketch, and cross-domain shifts, ELaTTA provides strong accuracy under strict and continual single-instance protocols. In the strict single-instance comparison, MEMO requires \emph{64.7

\times

} the computation and \emph{12.4

\times

} the peak memory of ELaTTA. A complementary keyword-spotting experiment shows that the same coefficient-search design transfers beyond vision. We further demonstrate ViT-Base adaptation on a Raspberry Pi 5, confirming the feasibility of deploying ELaTTA on an edge device.

Abstract

Test-Time Adaptation

Explore similar work

Efficient Test-Time Adaptation through Latent Subspace Coefficients Search

ZOTTA: Test-Time Adaptation with Gradient-Free Zeroth-Order Optimization

Cross-device Collaborative Test-time Adaptation with Zeroth-order Optimization and Model Merging