DeltaWAM: Change-Centric Visual Foresight via Delta Tokens for an Efficient World-Action Model

Tianyun Jiang; Wenrui Bao; Bingxin Xu; Yu Tian; Yuzhang Shang

cs.ROSep 27, 2026

DeltaWAM: Change-Centric Visual Foresight via Delta Tokens for an Efficient World-Action Model

Authors: Tianyun Jiang, Wenrui Bao, Bingxin Xu, Yu Tian, Yuzhang Shang

Organizations: Fudan University · University of Central Florida · University of Southern California

Abstract

World-Action Models (WAMs) offer visual foresight for robotic manipulation, but pixel-space models repeatedly reconstruct entire future scenes, incurring high computational cost and spatio-temporal redundancy. In physical manipulation, consecutive frames often share most of their visual context; the changes between them are what an action policy needs to anticipate. We introduce DeltaWAM, a change-centric WAM that makes a compact delta token the unit of future prediction. Each token is a single vector encoding changes between consecutive dense DINO feature maps. DeltaWAM builds on DeltaWorld, a latent world model pretrained on large-scale videos, to autoregressively predict one delta token per future frame. A flow-matching action expert then conditions on the predicted transitions and current DINO features, which serve as spatial anchors, to generate action chunks. Trained for 256 GPU hours on two H100 GPUs, DeltaWAM has 0.725B parameters and achieves 92.8% average success on LIBERO. It also shows robust generalization under procedural perturbations on LIBERO-Pro. Inference takes 142.1 ms per action chunk with 3.86 GB peak memory.

Figures & tables

Figure 1: DeltaWAM overview. (a) Existing policies represent an imagined future as a single subgoal or a sequence of dense frames or latent states, whereas DeltaWAM predicts one compact transition token per future step. (b) DeltaWAM couples a latent transition-centric world model DeltaWorld with an Action DiT, preserving efficient test-time visual foresight. (c) DeltaWAM is lightweight and has low latency compared with representative World-Action models.

Figure 2: Detailed DeltaWAM architecture. DeltaWAM couples the change-centric DeltaWorld with a flow-matching Action DiT. The pretrained DeltaWorld is first finetuned on demonstrations; then Action DiT is trained on robot data with the finetuned DeltaWorld frozen. DeltaWorld first rolls out predicted delta tokens autoregressively. Afterwards, both current DINO latents (the spatial anchor) and predicted delta tokens (the dynamic subgoal) are fed to Action DiT.

Figure 3: Language steers predicted futures toward the task goal. Wrist-view rollouts for LIBERO-Spatial task 9, which asks the robot to pick up the black bowl on the wooden cabinet and place it on the plate. In DeltaWorld’s unconditional predictions, the robot approaches the stove instead of the cabinet. This is misleading since the bowl is on the cabinet.

Figure 4: DeltaWAM training and inference. During the second training stage, Action DiT learns from oracle-selected or randomly sampled future trajectories produced by frozen DeltaWorld. At inference, one sampled future trajectory conditions action generation.

Method	Size	Emb. PT	Latency (ms)	Spatial	Object	Goal	Long	Average
Reactive Policies
Diffusion Policy [ 11 ]	263M	$\times$	–	78.3	92.5	68.3	50.5	72.4
OpenVLA [ 27 ]	7B	$\checkmark$	–	84.7	88.4	79.2	53.7	76.5
$\pi_{0}$ [ 6 ]	3.5B	$\checkmark$	220	98.0	96.8	94.4	88.4	94.4
$\pi_{0.5}$ [ 5 ]	3.5B	$\checkmark$	220	98.8	98.2	98.0	92.4	96.9
GR00T-N1.6 [ 37 ]	3.3B	$\checkmark$	259	97.7	98.5	97.5	94.4	97.0

Table 1: Standard LIBERO success rate (%), model size, and model-only latency per action-chunk query. Emb. PT denotes embodied action pretraining. Baseline sizes and latency values follow their source papers and the comparison protocol of LaWAM [ 10 ] .

Method	Spatial	Object	Goal	Long	Average
OpenVLA [ 27 ]	0.0	0.0	0.0	0.0	0.0
$\pi_{0}$ [ 6 ]	0.0	0.0	0.0	0.0	0.0
$\pi_{0.5}$ [ 5 ]	1.0	1.0	0.0	1.0	0.75
MolmoAct [ 28 ]	0.0	0.0	0.0	6.0	1.5
NORA [ 24 ]	0.0	0.0	0.0	0.0	0.0
x-VLA [ 55 ]	0.0	8.0	9.0	10.0	6.75

Table 2: Success rate (%) under the LIBERO-Pro task perturbation, broken down by the four standard task suites. Official benchmark baselines follow Zhou et al. [57] . We evaluated the results for Fast-WAM and Lingbot-VA ourselves.

Panel of Figure 5 — Figure 5: Delta-token structure under LIBERO-Pro task variations. The two t-SNE projections compare predicted delta tokens from a perturbed task (green triangles) with tokens from its nominal original task (blue) and an original task whose task requirements match the perturbation (orange).

Panel of Figure 6 — Figure 6: Training and Test Ablations. (a) Success rates of three models across the four task suites on LIBERO. Current DINO only, decoded future DINO, and delta tokens use 2,048, 14,336, and 2,060 visual tokens in total for one action chunk, respectively. (b) Comparison of success rates on LIBERO as the number of predicted future steps per camera view varies. Each step contributes one delta token per view, and the dashed line marks the default six-step horizon.

Figure 7: Real-world deployment We conduct real-world experiments on a Piper robotic arm. The task is to pick up the bottle/can and place it on the plate. DeltaWorld first predicts the delta tokens, then the Action DiT predicts the action chunk for execution. The frozen delta tokenizer decoder and DINO-to-RGB decoder is applied for DeltaWorld’s predictions.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	DeltaWorld adaptation	Action DiT training
Optimizer	AdamW	AdamW
Learning rate	$10^{-4}$	$5\times 10^{-5}$
Weight decay	$0.4$	$0.01$
Warmup steps	1,000	500
Batch size / GPU	32	2
Gradient accumulation	32	2

Appendix

Table 3: Training Hyperparameters.

Scope / Method	Camera	Robot	Light	Layout
OpenVLA [ 27 ]	0.8	3.5	8.1	28.5
NORA [ 24 ]	2.2	37.0	45.7	62.1
WorldVLA [ 9 ]	0.1	27.9	43.7	38.0
UniVLA [ 7 ]	1.8	46.2	69.0	31.9
$\pi_{0}$ [ 6 ]	13.8	6.0	85.0	68.9
$\pi_{0}$ -FAST [ 42 ]	65.1	21.6	73.2	68.8

Appendix

Table 4: Success rates (%) across selected LIBERO-Plus perturbation suites. Published baseline results are taken from the official benchmark leaderboard and the WAM robustness study [ 15 , 54 ] .

Figure 8: Delta-token structure under LIBERO-Plus lighting variation. LIBERO-Goal task embeddings under original light (circles) and a changed light condition (triangles).

Future steps	Spatial	Object	Goal	Long	Average
Default (6)	92.6	97.0	95.8	85.8	92.8
Full horizon (8)	88.8	95.4	94.6	81.2	90.0

Appendix

Table 5: LIBERO success rates (%) with six versus eight future steps.

Panel of Figure 11 — Figure 11: Task structure in predicted delta tokens across standard LIBERO suites. Each point concatenates synchronized agent-view and wrist-view tokens at one future step, yielding one joint embedding per step. Samples come from middle-phase windows of collected rollouts.

Panel of Figure 12 — Figure 12: Action DiT attention to predicted future delta tokens. Native cross-attention probabilities are averaged across Action DiT layers. Rows denote action positions within a generated chunk, columns denote predicted future steps, and agent and wrist tokens are shown separately. Probabilities remain normalized over the complete conditioning context, not only the displayed future tokens.

Explore similar work

DeltaWAM: Delta World Action Models for Bimanual Manipulation
Sep 23, 2026Han Yan, Zishang Xiang, Haokai Jiang +6Video Diffusion Models Bimanual Robotic Manipulation
School of Computer Science, Peking University · AI2 Robotics
DreamWAM: Beyond RGB Future Prediction for World Action Models
Aug 5, 2026Shanglin Yuan, Weiheng Zhao, Xin Shi +6World Model Learning Action-Conditioned World Models
Huazhong University of Science and Technology · D-Robotics · Wuhan University +1
Foresight Without Seeing: Latent Futures for World Action Models
Aug 12, 2026Jiakai Huang, Zhongbo Wu, Siyu Xu +5World Models for Robotics Visuomotor Policy Learning
Shanghai Jiao Tong University · ACE Robotics · Nanyang Technological University

Sep 23, 2026J/K move · Enter open · S save

DeltaWAM: Delta World Action Models for Bimanual Manipulation

Han Yan, Zishang Xiang, Haokai Jiang, Zeyu Zhang+5

School of Computer Science, Peking University · AI2 Robotics

Video Diffusion Models Bimanual Robotic Manipulation Action-Conditioned World Models Efficient Inference for World Action Models World Action Models

Open paper AskPDF arXiv

World-action models (WAMs) transfer visual and motion priors from pretrained video generators to robot control by jointly modeling visual dynamics and actions. Existing WAMs, however, predict dense future frames during training, repeatedly modeling largely unchanged content and coupling action-conditioned dynamics to nuisance appearance variations. At inference, processing each complete observation with the heavy video expert bottlenecks few-step action generation. Accordingly, we propose DeltaWAM, which jointly predicts visual deltas and actions using dense-anchor, sparse-delta, and action streams, with three architectures that differ in representation and computation sharing. We further develop Streaming Delta Memory (SDM), which updates cached anchor context with compact observed deltas, reducing heavy video-expert processing. On RoboTwin, DeltaWAM with SDM improves average success over Fast-WAM from 81.3% to 85.4% in the clean setting and from 75.8% to 83.9% under visual randomization. The three architectures reduce training FLOPs by 17.78-23.77%, while SDM reduces one-step inference latency and FLOPs by 36.57% and 31.55%, respectively; real-world evaluations further show the highest overall success rate and normalized progress among the evaluated policies. Code: https://github.com/AIGeeksGroup/DeltaWAM. Website: https://aigeeksgroup.github.io/DeltaWAM.

Abstract

Explore similar work

DeltaWAM: Delta World Action Models for Bimanual Manipulation

DreamWAM: Beyond RGB Future Prediction for World Action Models

Foresight Without Seeing: Latent Futures for World Action Models