PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models

Yude Zou; Yifei Yao; Guanbo Jiao; Jianming Ma; Guanjie Zheng; Xing Gao

cs.CVOct 19, 2026

PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models

Authors: Yude Zou, Yifei Yao, Guanbo Jiao, Jianming Ma, Guanjie Zheng, Xing Gao

Organizations: Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · Fudan University

Abstract

World Action Models (WAMs) jointly predict future video and generate actions, yet remain sensitive to task-irrelevant visual perturbations, i.e., changes in visual appearance that should not affect robot behavior. We identify a gap in conventional WAM training: video prediction encourages representations to preserve appearance information, while action generation does not explicitly enforce invariance to task-irrelevant appearance. Data augmentation exposes WAMs to such perturbations but does not require perturbed and nominal observations to produce consistent action representations, and alignment with external pretrained features does not guarantee the desired invariance. To address this gap, we propose PICA-WAM, a self-supervised training framework that casts perturbation invariance as an auxiliary cross-appearance self-alignment task for WAMs. An exponential-moving-average (EMA) teacher processes the nominal observation to provide a deeper-layer target, while the online WAM processes a perturbed observation and aligns its shallower-layer action representation to this target through a lightweight projection head. Across four strong WAMs evaluated on RoboTwin 2.0, LIBERO-Plus, EBench, and a real robot, PICA-WAM consistently improves robustness under visual perturbations, preserves nominal performance in simulation, and raises the real robot success rate. PICA-WAM further reduces the sensitivity of video features to task-irrelevant appearance and its propagation into actions, as shown by paired layer-wise diagnostics. Code: https://github.com/YudeZou/PICA-WAM-release

Figures & tables

Figure 1: Overview of PICA-WAM . Top: augmentation supervises each observation independently, whereas PICA-WAM aligns shallower-layer action representations of the student on the perturbed observation to deeper-layer ones of its EMA teacher on the nominal observation. Bottom left: schematic of action representations. Matching shapes denote the same underlying task state across nominal and perturbed observations. Bottom right: performance. Ours is the best PICA-WAM result, and each axis is scaled linearly to the range of results per benchmark.

Figure 2: Framework of PICA-WAM . (a) A two-stream WAM is trained with $\mathcal{L}_{\text{gen}}=\mathcal{L}_{\text{video}}+\mathcal{L}_{\text{action}}$ . PICA-WAM adds an auxiliary self-alignment task between the action streams of the model and its EMA teacher. (b) The student $\theta$ reads a perturbed observation at video noise level $\sigma$ and is trained with $\mathcal{L}_{\text{gen}}$ . The EMA teacher $\phi$ reads the nominal observation of the same underlying task state at $\sigma-\delta$ . Projected shallower-layer action representations of the student are aligned to deeper-layer ones of the teacher under stop-gradient (sg), giving $\mathcal{L}_{\text{align}}$ .

Figure 3: Video-stream appearance sensitivity and its transfer into actions. Left: layer-wise sensitivity of video features on unseen perturbation families (Section 4.2 ). Right: action deviation caused by swapping the perturbed video features of one layer band into a nominal prediction, as a median ratio to that of the Baseline on the same observation pairs. In the highlighted layers 21–25, the sensitivity of +Aug rebounds, and these layers retain 84% of the action leakage of the Baseline under +Aug, compared with 35% under PICA-WAM .

	Video Sensitivity $\downarrow$		Action Sensitivity $\downarrow$		Drift (mm) $\downarrow$
Method	Seen	Unseen	Seen	Unseen	Seen	Unseen
Baseline	0.610	0.755	0.407	0.486	16.14	16.42
+ Aug	0.530	0.655	0.146	0.189	9.41	11.24
PICA-WAM (Ours)	0.369	0.455	0.057	0.075	5.49	6.21

Table 1: Paired Consistency under LIBERO-Plus Visual Perturbations. Sensitivity is measured at layer 8 of each stream, and drift is the end-effector distance after executing the 32-step action chunks predicted from the paired observations. Lower is better. Metrics and seen/unseen families are defined in Section 4.2 . Bold and underline denote the best and second-best values.

Method without Emb. PT.
Method	Clean	Random	Avg.
StarVLA ( Ye et al., 2026a )	46.5	3.2	24.8
AHA-WAM ( Cai et al., 2026a )	64.3	3.2	33.8
Fast-WAM ( Yuan et al., 2026b )	77.8	1.9	39.9
Copper-Policy ( Feng et al., 2026 )	70.8	13.0	41.9
CSWAM † ( Liu et al., 2026a )	78.5	9.4	43.9

Table 2: Evaluation Results on RoboTwin 2.0 Clean-to-Random. Method groups distinguish models with and without large-scale embodied pretraining. † , ‡ , and § denote methods built on the Fast-WAM, LingBot-VA, and InternW0-

\Delta

backbones, respectively. Bold and underline mark the best and second-best values within each group.

Method without Emb. PT.
		LIBERO-Plus
Method	LIBERO	Robot	Language	Light	BG	Noise	Layout	Camera	Avg.
Fast-WAM ( Yuan et al., 2026b )	97.6	44.5	68.9	78.2	53.7	37.7	60.7	16.4	51.5
Robust-WAM † ( Yan et al., 2026 )	97.9	43.4	73.6	90.2	62.6	56.8	67.8	28.7	58.9
4D-WAM † ( Yang et al., 2026b )	98.6	64.3	90.6	94.3	57.7	69.1	79.2	45.2	71.0
GaussianWAM † ( Zhang et al., 2026b )	97.6	61.6	75.4	90.8	58.2	83.0	76.7	54.7	71.3

Table 3: Evaluation Results on LIBERO and LIBERO-Plus. Method groups distinguish models with and without large-scale embodied pretraining before LIBERO adaptation. † and § denote methods built on the Fast-WAM and InternW0-

\Delta

backbones, respectively. Bold denotes the best value and underline the second best within each group.

	Mobile Manipulation		Table-Top Dex. & Prec.		Overall
Method	SR	Score	SR	Score	SR	Score
X-VLA ( Zheng et al., 2026 )	29.3	40.2	8.6	24.1	23.7	35.8
Fast-WAM ( Yuan et al., 2026b )	34.0	46.1	2.9	13.0	25.6	37.1
Cosmos3-Edge ( Agarwal et al., 2026 )	33.2	46.6	18.6	31.6	29.3	42.5
GigaBrain-0.7 ( Team et al., 2026 )	31.6	41.4	37.9	59.0	33.3	46.0
$\pi_{0}$ ( Black et al., 2024 )	34.0	46.8	32.9	49.4	33.7	47.5

Table 4: Evaluation Results on EBench . Mobile Manipulation contains 19 tasks and Table-Top Dexterous-and-Precise contains 7 tasks, and all averages weight tasks equally. † denotes a method built on the OpenWAM-

\alpha

backbone. Results of other methods are taken from the EBench leaderboard. Bold and underline mark the best and second-best values.

Method	Light	BG	Noise	Layout	Camera	Avg.
Baseline	93.2	60.6	67.2	81.4	45.7	69.6
+ SRA ( Jiang et al., 2025a )	94.1	60.6	70.1	82.5	50.8	71.6
+ Aug	93.3	74.2	87.1	78.3	62.1	79.0
+ SRA	92.7	80.0	92.1	79.0	62.7	81.3
+ PICA-WAM align video	94.0	79.0	92.8	82.2	60.0	81.6
+ PICA-WAM align both	95.4	76.3	93.6	79.8	58.7	80.7

Table 5: Ablation Results on LIBERO-Plus. We report success rates (%) and their unweighted mean on the five dimensions that perturb the visual scene. Robot initial state and language are excluded because they perturb the initial pose and the instruction. Baseline and +Aug are the controlled variants of Section 4.1 . + SRA adds self-alignment with the noise-level and layer gaps but without the appearance gap. The PICA-WAM rows add the alignment of Eq. ( 3 ) to +Aug and differ in the aligned stream. Bold and underline denote the best and second-best values.

Figure 4: Real-world evaluation scene under the four lighting conditions. The photos show the eight task objects on the evaluation tabletop, which differs from the one used for the demonstrations. The bowl, the mouse, and the pen are new instances. The demonstrations are recorded under natural lighting, whereas the blue, green, and orange lights shift the apparent colors of the tabletop and the objects. For example, the banana appears pink under blue light.

Method	Banana	Bowl ∗	Cucumber	Mouse ∗	Orange	Pen ∗	Cube	Tissue	Avg.
$\pi_{0.5}$	85.0	100.0	85.0	0.0	0.0	100.0	0.0	75.0	55.6
OpenWAM- $\alpha$	5.0	95.0	5.0	25.0	0.0	70.0	0.0	65.0	33.1
PICA-WAM (Ours) †	95.0	100.0	90.0	30.0	75.0	100.0	10.0	40.0	67.5

Table 6: Real-World Success Rates per Task (%). Each task is evaluated in 20 trials, five under each lighting condition. Avg. is over all 160 trials. ∗ marks objects replaced with new instances at evaluation. † denotes a method built on the OpenWAM-

\alpha

backbone. Bold and underline mark the best and second-best values.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Aligned stream	action
Student layer $m$ / teacher layer $n$	8 / 20
Loss weight $\lambda$	0.05
Noise-level gap $\delta$	$\mathcal{U}[0,0.2)$
Projection head	Linear–GELU–Linear, hidden width 512
EMA decay	0.9999 (fp32)

Appendix

Table 7: Alignment hyperparameters shared by Fast-WAM, LingBot-VA, InternW0-

\Delta

, and OpenWAM-

\alpha

. Layers are indexed from 0.

Benchmark	Model	Steps	Batch
RoboTwin 2.0	LingBot-VA	30,000	16
	Fast-WAM	21,145	128
	InternW0- $\Delta$	5,370	1,024
LIBERO	Fast-WAM	21,700	128
LIBERO	InternW0- $\Delta$	16,275	256
EBench	OpenWAM- $\alpha$	100,000	256

Appendix

Table 8: Training configurations. InternW0-

\Delta

and OpenWAM-

\alpha

are trained on 33-frame clips, at a stride of 4 from three cameras for OpenWAM-

\alpha

. LingBot-VA’s batch size counts episodes.

Figure 5: Executed trajectories under visual perturbations. Each column shows one LIBERO-Plus state under a perturbed observation. Each curve is the end-effector path obtained by executing the 32-step action chunk predicted from a different perturbed observation of the same state with the same noise seed. The white dot marks the start position. A tighter fan indicates more consistent actions across appearances. The first three columns are the states with the largest baseline drift. The last two are drawn at random.

	Nominal observation		Perturbed observation
Method	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
Baseline	24.62	0.864	19.82	0.734
+ Aug	22.38	0.829	24.14	0.813
PICA-WAM	22.44	0.828	24.12	0.813

Appendix

Table 9: Video prediction fidelity on 96 LIBERO training clips, averaged over the third-person and wrist cameras, with 95% bootstrap confidence intervals. The perturbed observations are constructed with the training distribution

T

World Action Models

World Action Models connect predictions of how scenes change under action to executable policies for robots, vehicles, and other embodied agents.

See all 223 papers

ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?
Jun 17, 2026Yuyang Zhang, Wenyao Zhang, Zekun Qi +7Action-Conditioned World Models World Action Models
Shanghai Jiao Tong University · Tsinghua University · Tencent Robotics X +1
What Visual Representations Do World Action Models Need?
Oct 9, 2026Kai Liu, Jiangran Lyu, Tingyu Deng +9World Action Models World Models for Robotics
Institute of Automation, Chinese Academy of Sciences · Galbot · Peking University +2
$R^2$ -WAM: Repair-and-Reject Post-Training for World Action Models
Oct 4, 2026Ruiyan Xu, Haisheng Su, Sixu Lin +4World Action Models Action-Conditioned World Models
The Chinese University of Hong Kong, Shenzhen · Manifold AI · Southeast University +1

Jun 17, 2026J/K move · Enter open · S save

ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?

Yuyang Zhang, Wenyao Zhang, Zekun Qi, He Zhang+6

Shanghai Jiao Tong University · Tsinghua University · Tencent Robotics X+1

Action-Conditioned World Models World Action Models Embodied World Models

Open paper AskPDF arXiv

World Action Models (WAMs) commonly rely on video generation to bridge visual world modeling and robot control. However, video-based WAMs face three coupled limitations: dense multi-frame future tokens make inference costly, full video prediction spends capacity on action-irrelevant temporal and appearance details, and long-horizon future imagination may introduce errors that mislead action prediction. These issues raise a simple question: Does world action model really need video generation? We propose ImageWAM, a simple WAM framework that repurposes pretrained image editing models for robot action prediction. In contrast to video generation, image editing provides a better-matched prior: it only needs to model a target-frame transformation, focuses on action-relevant current-to-target visual differences, and grounds task instructions to localized visual changes through edit pretraining. In practice, ImageWAM does not decode the target frame at inference time; instead, it conditions a flow-matching action expert on the KV caches produced by image-editing denoising, using them as a compact world-action context. ImageWAM outperforms standard VLA baselines and matching competitive WAMs without additional policy pretraining across different simulator and real-world experiments. It also reduces FLOPs to 1/6 and latency to 1/4 of video-based WAMs. Attention analysis further shows that editing caches focus on task-relevant change regions, supporting image editing as an effective alternative to video-based world-action modeling.

Abstract

World Action Models

Explore similar work

ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?

What Visual Representations Do World Action Models Need?

R2R^2R2-WAM: Repair-and-Reject Post-Training for World Action Models

$R^2$ -WAM: Repair-and-Reject Post-Training for World Action Models