SLIP-VLA: Single-Step Latent Imagination for Policy Learning in Vision-Language-Action Models

Tianfu Li; Haoxuan Xu; Wenbo Chen; Haitian Li; Changchuan Yang; Xinhu Zheng; Jun Ma; Yuan Liu

cs.ROSep 27, 2026

SLIP-VLA: Single-Step Latent Imagination for Policy Learning in Vision-Language-Action Models

Authors: Tianfu Li, Haoxuan Xu, Wenbo Chen, Haitian Li, Changchuan Yang, Xinhu Zheng, Jun Ma, Yuan Liu, +2 more

Organizations: The Hong Kong University of Science and Technology (Guangzhou). · The Hong Kong University of Science and Technology. · Nanyang Technological University. · Zhejiang University.

Abstract

Vision-Language-Action models are increasingly effective for robotic manipulation, yet most predict actions directly from current observations without explicitly modeling future scene evolution. Recent methods introduce future prediction to improve action generation, but dense future modeling often requires expensive iterative denoising, while one-step alternatives can underperform their multi-step counterparts. To reconcile efficient future modeling with strong action performance, we present SLIP-VLA, a policy learning framework that equips VLA models with a Single-Step Latent Imagination for future-aware action prediction. SLIP-VLA obtains temporally dense future latent representations with a single denoising update, and we improve the perceptual sufficiency of these representations by aligning intermediate latents with future geometric and semantic features. We further improve their control sufficiency through action-conditioned latent world modeling and inverse dynamics modeling, explicitly coupling latent transitions with robot actions. SLIP-VLA achieves state-of-the-art performance across diverse simulation benchmarks and real-world manipulation tasks, while its single-step latent imagination takes only 12 ms.

Figures & tables

Fig. 2: Pipeline of SLIP-VLA. The VLM backbone first encodes the multimodal context, and the World DiT then performs a single denoising pass to produce temporally dense future latents. (a) For perceptual sufficiency shaping, selected intermediate representations are aligned with future VGGT- $\Omega$ geometric features and DINOv3 semantic features. (b) For control sufficiency shaping, deeper latent representations are regularized through forward and inverse latent dynamics to preserve action-dependent transitions. The shaped future representations are injected layer-wise into the Action DiT for action generation, while all auxiliary shaping modules are used only during training.

Method	Spatial	Object	Goal	Long	Avg.
General VLAs
OpenVLA [ 1 ]	84.7	88.4	79.2	53.7	76.5
OpenVLA-OFT [ 35 ]	97.6	98.4	97.9	94.5	97.1
$\pi_{0}$ [ 2 ]	98.0	96.8	94.4	88.4	94.4
$\pi_{0.5}$ [ 3 ]	98.8	98.2	98.0	92.4	96.9
Future-Aware VLAs

TABLE I: LIBERO success rates (%). Best and second-best results are bolded and underlined , respectively.

Method	Cam.	Robot	Lang.	Light	BG	Noise	Layout	Avg.
General VLAs
OpenVLA [ 1 ]	0.8	3.5	23.0	8.1	34.8	15.2	28.5	15.6
OpenVLA-OFT [ 35 ]	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
$\pi_{0}$ [ 2 ]	13.8	6.0	58.8	85.0	81.4	79.0	68.9	53.6
$\pi_{0}$ -Fast [ 36 ]	65.1	21.6	61.0	73.2	73.2	74.4	68.8	61.6
Future-Aware VLAs

TABLE II: LIBERO-Plus success rates (%). Best and second-best results are bolded and underlined , respectively.

Method	Clean	Rand.	Avg.
General VLAs
$\pi_{0}$ [ 2 ]	65.9	58.4	62.2
$\pi_{0.5}$ [ 3 ]	82.7	76.8	79.8
Future-Aware VLAs
WorldVLA [ 4 ]	42.5	32.2	37.4
World Action Models

TABLE III: RoboTwin 2.0 success rates (%). Best and second-best results are bolded and underlined , respectively.

Fig. 3: Qualitative analysis of perceptual and control information in the single-step latent. (a) Compared with the naive single-step baseline, SLIP-VLA recovers more coherent future geometry and more concentrated task-relevant semantic responses. (b) For forward dynamics, the correct action produces a prediction closer to the target latent, while a shuffled action induces clear spatial deviations; the action-response map highlights regions most affected by the action change.

Fig. 4: Effect of denoising steps on future prediction and control performance. The top row shows representative predicted videos with one to four denoising steps, while the curves report PSNR and LIBERO-Plus success rate. Additional denoising provides no consistent improvement in visual prediction quality and does not improve action performance, with the highest success rate achieved using a single step.

Fig. 5: Real-world evaluation of SLIP-VLA. (a) The real-world manipulation platform. (b.1) Our SLIP-VLA successfully completes the in-distribution corn-to-plate task. (b.2) It remains successful under visual distribution shifts in background, object appearance, and object orientation. (b.3) SLIP-VLA further completes both two-step and three-step long-horizon tasks involving sequential interaction with the pot and the target object.

Variant	LIBERO	LIBERO-Plus
Naive Single-Step	97.7	79.0
w/ Perceptual Sufficiency	98.3	81.4
w/ Control Sufficiency	98.5	82.3
w/ Perceptual + Control Sufficiency	98.9	83.2

TABLE IV: Ablation study of perceptual and control sufficiency shaping. Success rates (%) on LIBERO and LIBERO-Plus.

	Geometry		Semantics
Variant	Pearson $\uparrow$	AbsRel $\downarrow$	mIoU $\uparrow$	Dice $\uparrow$
Naive Single-Step	0.804	0.188	0.388	0.486
SLIP-VLA (Ours)	0.916	0.112	0.590	0.683

TABLE V: Quantitative analysis of the perceptual sufficiency of frozen single-step latent representations.

Evaluation Condition	Distance $\downarrow$	$\Delta$ vs. Correct $\uparrow$
Forward Dynamics
Correct Action	0.012	–
Gripper Shuffled	0.527	0.515
Motion Shuffled	0.772	0.759
Fully Shuffled	0.876	0.864
Inverse Dynamics

TABLE VI: Quantitative analysis of control sufficiency in SLIP-VLA. Forward dynamics measures latent prediction distance under different action perturbations, while inverse dynamics measures action recovery error under different latent-transition perturbations. Lower is better.

Source Initialization	LIBERO	LIBERO-Plus
Gaussian Noise	98.6	82.5
Current Latent	98.8	82.9
Current + Noise (Ours)	98.9	83.2

TABLE VII: Ablation study of source initialization strategies. Success rates (%) on LIBERO and LIBERO-Plus.

Task	$\pi_{0}$	SLIP-VLA (Ours)
In-Distribution
Corn $\rightarrow$ Plate	72	86
Out-of-Distribution
Unseen Background	58	70
Unseen Object Appearance	54	70
Unseen Object Orientation	52	66

TABLE VIII: Real-world success rates (%). Each task condition is evaluated over 50 trials.

Explore similar work

SLIM-0.5B: Learning Action-Grounded Predictive Latents for Robot Manipulation
Aug 10, 2026Jingkai Wang, Zihan Tang, Gu Zhang +7Goal-Conditioned Dynamic Manipulation Robotic Manipulation
1Fudan University · 2Beijing Academy of Artificial Intelligence · 3Tsinghua University +1
Imagine-RL: Residual-Confidence-Guided Cross-Attention for World-Model-Augmented VLA Reinforcement Learning
Sep 21, 2026Kejia Hu, Wentong Zhai, Bo Zhao +1Diffusion-Based Vision-Language-Actions Imagination
Department of Electric Engineering, Korea Advanced Institute of Science and Technology · University Of Science and Technology Beijing · Shanghai Jiao Tong University +1
Being-H0.7: A Latent World-Action Model from Egocentric Videos
Apr 30, 2026Hao Luo, Wanpeng Zhang, Yicheng Feng +6Latent World Models Sparse Supervision
BeingBeyond Team

Aug 10, 2026, cs.ROJ/K move · Enter open · S save

SLIM-0.5B: Learning Action-Grounded Predictive Latents for Robot Manipulation

Jingkai Wang, Zihan Tang, Gu Zhang, Mingyu Cao+6

1Fudan University · 2Beijing Academy of Artificial Intelligence · 3Tsinghua University+1

Goal-Conditioned Dynamic Manipulation Robotic Manipulation Latent Prediction Robot Systems

Open paper AskPDF arXiv

Vision-language-action policies rely on large multimodal backbones to jointly perform perception, language conditioning, and action generation at every control step. Much of this capacity supports open-domain semantics, whereas continuous robot manipulation primarily requires compact representations of observations, actions, and the transitions induced by actions. Pixel-level world models provide another route, but predicting visual details irrelevant to control can be unnecessarily expensive. We propose SLIM (Self-supervised Latent Interaction Model), a compact 0.5B-parameter latent interaction policy. SLIM learns action-grounded predictive latents that capture both action-conditioned future transitions and the actions that explain observed changes. SLIM learns these representations through self-supervised masked trajectory prediction, combining action reconstruction with future-latent prediction. A compact Mixture-of-Transformers (MoT) backbone models interactions between observation latents and action tokens. The resulting policy is trained with flow matching for language-conditioned action generation. Across simulation benchmarks and real-world evaluation, SLIM matches or exceeds representative large-scale VLA and world-action-model baselines with fewer parameters, no additional embodied pretraining, lower inference latency, and substantially lower GPU memory usage.

Abstract

Explore similar work

SLIM-0.5B: Learning Action-Grounded Predictive Latents for Robot Manipulation

Imagine-RL: Residual-Confidence-Guided Cross-Attention for World-Model-Augmented VLA Reinforcement Learning

Being-H0.7: A Latent World-Action Model from Egocentric Videos