Literature
Display
← All papers

Fast-WAM: Do World Action Models Need Test-time Future Imagination?

Tianyuan Yuan · Zibin Dong · Yicheng Liu · Hang Zhao

Preprint posted · NeurIPS 2026 (accepted) ↗

Robotics8 min text · CondensedOriginal paper ↗
Contents

At a glance

Fast-WAM uses future-video prediction as a training objective but omits future-video generation during deployment. A pretrained video backbone encodes the current observation once, and an action expert denoises an action chunk. Controlled variants suggest that video co-training matters more than explicit test-time imagination on the evaluated tasks. Reported inference latency is 190 ms, versus 580 ms for joint generation and 810 ms for sequential video-then-action generation.

1 Introduction

A world action model can benefit from predicting video in two different ways: the training loss can improve its representation, and generated future observations can guide actions at inference. Existing systems often combine both, obscuring which accounts for improved control. Fast-WAM separates them using a shared implementation with joint-generation, sequential-generation and no-video-co-training controls.

The central result is empirical: retaining video supervision while skipping future synthesis remains competitive on LIBERO, RoboTwin and one real towel-folding task. This does not prove that future imagination is unnecessary for all planning or out-of-distribution control. The comparison omits outer autoregressive rollout and tests a particular model/data regime. This edition covers all 13 pages of v2, dated 23 March 2026, including four complete numbered figures, both Figure 2 panels, all three tables, nine equations and 38 references.

Complete original PDF S1.F1
Figure 1: Three representative WAM paradigms. (A) Joint-modeling WAMs denoise future video and action tokens together. (B) Causal WAMs first generate future observations and then condition action prediction on the generated future representation. (C) Fast-WAM retains video co-training during training but removes explicit future generation at inference time, directly predicting actions from latent world representations in a single forward pass.

2 Related Work

Vision-Language-Action Policies.

Policies such as OpenVLA and π₀.₅ inherit semantic knowledge from pretrained vision-language components and directly generate actions. Fast-WAM keeps a similarly direct deployment interface but uses future-video supervision to adapt its representation. This is a distinction between objectives and inference procedures, rather than a claim that every VLA lacks temporal knowledge.

World Action Models and Video-based Robot Policies.

Video-policy methods include joint video/action denoising and video-first action prediction, represented here by work such as Motus and LingBot-VA. Prior Video Prediction Policy and Unified Video Action Model also exploit predictive representations or reduce decoding cost. Fast-WAM emphasizes controlled comparisons of the training loss and inference dependency rather than claiming to originate every form of implicit video-based control.

3 Method

3.1 Problem Formulation

The standard policy predicts an action chunk from the current observation and instruction. An imagine-then-execute formulation introduces a future-video variable and integrates the action distribution conditioned on that future. Fast-WAM instead conditions actions on a deterministic latent representation z(o,l), obtained from the current context in one video-backbone pass.

The future-prediction loss still trains the backbone. Removing sampled future frames does not remove the pretrained video prior or all internal representations of dynamics. The experiment isolates explicit synthesis in these implementations, not whether the model internally represents future consequences.

Equations 1–4
p(a1:H∣o,l),p(a_{1:H}\mid o,l), (1)
p(a1:H∣o,l)=∫p(v1:T∣o,l)p(a1:H∣o,l,v1:T)dv1:T,p(a_{1:H}\mid o,l)=\int p(v_{1:T}\mid o,l)\,p(a_{1:H}\mid o,l,v_{1:T})\,dv_{1:T}, (2)
pθ(a1:H∣o,l),p_{\theta}(a_{1:H}\mid o,l), (3)
pθ(a1:H∣o,l)=pθ(a1:H∣z(o,l)).p_{\theta}(a_{1:H}\mid o,l)=p_{\theta}(a_{1:H}\mid z(o,l)). (4)

3.2 Model Architecture

Overview.

Training includes clean first-frame tokens, noisy future-video tokens and noisy action tokens. At deployment, the future tokens disappear. The clean visual context is encoded once and supplies features for action generation. “Single pass” refers to the video encoder: the action expert still performs ten denoising steps.

Architecture.

The backbone is a Wan video diffusion transformer, instantiated as Wan2.2-5B, with its pretrained T5 text encoder and video VAE. A separate action transformer shares attention with the video branch. Language reaches both branches through cross-attention.

The mask is the key design: future-video tokens can access each other and the clean first frame; actions can access each other and the clean first frame, but cannot read future-video tokens. Clean first-frame tokens do not read future/action tokens. Thus video targets cannot leak into action conditioning during training. At inference, the current-frame representation can be cached while the action expert denoises. Figure 2 retains the complete architecture and both training/inference masks.

Complete original PDF S3.F2.sf1
(a) Fast-WAM model architecture.
Complete original PDF S3.F2.sf2
(b) Training and inference masks.
Figure 2: Fast-WAM architecture and the structured attention mask used to disentangle video co-training from action generation.

Training objective.

Both branches use flow matching: interpolate a clean target with Gaussian noise and regress the noise-minus-target velocity field. The total loss is action loss plus λ times the future-video loss. Action targets are chunks; video targets are VAE latents. The paper gives the objective but does not specify λ or every training/batch/image-resolution choice needed for a full reproduction.

Equations 5–9
yt=(1−t)​y+t​ϵ.y_{t}=(1-t)\,y+t\,\epsilon. (5)
ℒFM​(y)=𝔼y,ϵ,t​[‖fθ​(yt,t,o,l)−(ϵ−y)‖22].\mathcal{L}_{\mathrm{FM}}(y)=\mathbb{E}_{y,\,\epsilon,\,t}\left[\left\|f_{\theta}(y_{t},\,t,\,o,\,l)-(\epsilon-y)\right\|_{2}^{2}\right]. (6)
ℒact=ℒFM(a1:H).\mathcal{L}_{\mathrm{act}}=\mathcal{L}_{\mathrm{FM}}(a_{1:H}). (7)
ℒvid=ℒFM(z1:T).\mathcal{L}_{\mathrm{vid}}=\mathcal{L}_{\mathrm{FM}}(z_{1:T}). (8)
ℒ=ℒact+λ​ℒvid,\mathcal{L}=\mathcal{L}_{\mathrm{act}}+\lambda\mathcal{L}_{\mathrm{vid}}, (9)

3.3 Controlled Variants for Disentangled WAM Design

Fast-WAM-Joint allows video/action attention and denoises both together. Fast-WAM-IDM first generates future video, then conditions the action expert on that representation; its ground-truth video inputs receive noise augmentation with probability 0.5 during training. The no-video-co-training variant keeps Fast-WAM’s architecture and deployment procedure while removing the auxiliary video objective.

This last ablation is particularly informative because its inference latency stays unchanged. However, it still starts from a pretrained video backbone: the experiment removes robot-task video co-training, not all video learning. The imagination variants also differ in attention and training choices, so the comparison supports a design-level tradeoff rather than isolating a single mathematical variable under identical optimization.

4 Experiment

4.1 Implementation Details

The 5B video backbone and 1B action expert total approximately 6B parameters; the action expert uses hidden width 1,024. Actions have horizon 32, and fourfold temporal subsampling gives nine video frames per chunk. Multiple camera views are concatenated before VAE encoding.

Inference uses ten action-denoising steps and classifier-free guidance scale 1.0. The schedule is described as logit-normal. Training uses AdamW at 10⁻⁴ learning rate, 0.01 weight decay, cosine decay, mixed precision and gradient clipping at 1.0. Reported latency is measured on one RTX 5090D V2 with 32 GB memory. “Without embodied pretraining” means no additional robot-data pretraining, not training from scratch or absence of large-scale video/text pretraining.

4.2 Experiment Setup

The evaluation combines two simulation benchmarks with a single physical towel-folding task. Success percentages summarize task outcomes; inference latency measures computational delay, while physical completion time includes the behavior selected by the policy. These are distinct metrics.

LIBERO.

LIBERO comprises four suites of ten tasks, each with 500 demonstrations. Models train for 20,000 steps. Evaluation uses 2,000 trials across 40 tasks with varied random seeds. The paper does not report uncertainty across independent training runs, so small differences near saturation are not established statistical equivalence.

RoboTwin 2.0.

RoboTwin 2.0 training mixes 2,500 clean and 25,000 randomized demonstrations, with 30,000 optimization steps. Evaluation reports 100 trials per task under each clean/randomized condition. Although the prose says “over 50” tasks, the complete results table contains exactly 50 task rows. Its 14 column means recompute exactly to the printed averages.

Complete original PDF S4.F3
Figure 3: Real-world towel-folding task on the Galaxea R1 Lite platform. Folding a deformable object requires long-horizon planning and precise closed-loop manipulation, making it a challenging benchmark for evaluating both task success and execution efficiency.

Real-World Evaluation.

The real experiment uses 60 hours of teleoperated towel folding on Galaxea R1 Lite, with 30,000 training steps. It measures success and mean completion time to distinguish an efficient fold from repeated correction. Figure 3 retains all eight demonstration images. Trial counts, success criteria, timeout/censoring rules and confidence intervals are not specified, limiting interpretation of the scatter plot.

4.3 Main Results

4.3.1 Overall comparison on simulation benchmarks

Fast-WAM scores 91.88% clean and 91.78% randomized on RoboTwin, averaging 91.8%. Pretrained LingBot-VA is slightly higher at 92.2%; pretrained Motus scores 87.8%, and Motus initialized from Wan without embodied pretraining scores 77.3%. The 80.6% LingBot-VA-from-Wan comparison has only a clean result, with the randomized cell missing; it is not a matched two-condition average.

On LIBERO, Fast-WAM gives Spatial/Object/Goal/Long scores of 98.2/100.0/97.0/95.2%, averaging 97.6%. The reported pretrained baselines include π₀.₅ at 96.9%, Motus at 97.7% and LingBot-VA at 98.5%. These are strong benchmark results, but the table mixes methods with different pretraining data and does not make Fast-WAM uniformly best.

Table 1: Results on RoboTwin. Fast-WAM matches strong pretrained WAM baselines without using embodied pretraining, while the two imagine-then-execute variants remain highly comparable and removing video co-training causes a substantial drop.
Method Embodied PT. Clean Rand. Average
π0\pi_{0} [10] ✓ 65.92 58.40 62.2
π0.5\pi_{0.5} [11] ✓ 82.74 76.76 79.8
Motus [5] ✓ 88.66 87.02 87.8
Motus from WAN2.2 ✗ 77.56 77.00 77.3
LingBot-VA [3] ✓ 92.90 91.50 92.2
LingBot-VA from WAN2.2 ✗ 80.60 – 80.6
Fast-WAM (Ours) ✗ 91.88 91.78 91.8
Fast-WAM Variants
Fast-WAM ✗ 91.88 91.78 91.8
Fast-WAM-Joint ✗ 90.84 90.32 90.6
Fast-WAM-IDM ✗ 91.16 91.34 91.3
Fast-WAM w.o. video co-train ✗ 82.76 84.80 83.8
Table 2: Results on LIBERO. Fast-WAM achieves competitive overall performance without embodied pretraining, remains close to both imagine-then-execute variants, and outperforms the no-video-co-training ablation by a clear margin.
Method Embodied PT. Spatial Object Goal Long Average
OpenVLA [9] ✓ 84.7 88.4 79.2 53.7 76.5
π0\pi_{0} [10] ✓ 96.8 98.8 95.8 85.2 94.1
π0.5\pi_{0.5} [11] ✓ 98.8 98.2 98.0 92.4 96.9
LingBot-VA [3] ✓ 98.5 99.6 97.2 98.5 98.5
Motus [5] ✓ 96.8 99.8 96.6 97.6 97.7
Fast-WAM (Ours) ✗ 98.2 100.0 97.0 95.2 97.6
Fast-WAM Variants
Fast-WAM ✗ 98.2 100.0 97.0 95.2 97.6
Fast-WAM-Joint ✗ 99.6 99.4 98.2 96.8 98.5
Fast-WAM-IDM ✗ 98.8 97.8 97.8 97.6 98.0
Fast-WAM w.o. video co-train ✗ 89.2 99.2 95.4 90.0 93.5

4.3.2 Controlled comparison with Fast-WAM variants

On RoboTwin, Fast-WAM/Joint/IDM score 91.8/90.6/91.3%; removing video co-training gives 83.8%, an 8.0-point drop. On LIBERO, the corresponding scores are 97.6/98.5/98.0%, versus 93.5% without video co-training, a 4.1-point drop. The largest LIBERO ablation losses are Spatial (98.2→89.2%) and Long (95.2→90.0%).

These differences support the auxiliary training objective in this setting. Explicit imagination is not consistently better or worse across the two benchmarks, and no confidence intervals establish equivalence between close scores. Better task performance is compatible with improved representations, but the paper does not directly measure their physical accuracy or causal content.

4.3.3 Real-world performance and efficiency

Pretrained π₀.₅ has the best plotted real-world success and completion time. Within the Fast-WAM family, IDM has the highest success, while Fast-WAM finishes faster. The plot places Fast-WAM at roughly 75% success and 150 seconds, IDM near 90% and 175 seconds, and Joint near 70% and 225 seconds; these are visual readings, not tabulated measurements. The no-video variant is explicitly reported at only 10% success and has the longest completion time.

Latencies are 180 ms for π₀.₅, 190 ms for Fast-WAM and its no-video ablation, 580 ms for Joint, and 810 ms for IDM. Thus Fast-WAM is approximately 3.1× faster than Joint and 4.3× faster than IDM; the headline “over 4×” applies to the latter comparison. It is not faster than π₀.₅ in this chart. Chunk inference latency is not the per-action actuator update period, and physical completion time cannot be inferred from latency alone.

The fixed-latency auxiliary-loss ablation provides stronger evidence for training benefit than the plot provides for universal dispensability of imagination. One real task, unspecified rollout counts and no uncertainty estimates leave the generalization scope open.

Complete original PDF S4.F4
Figure 4: Real-world results on the long-horizon towel-folding task. The left panel plots success rate against average completion time, where upper-left is better. The right panel compares inference latency. Fast-WAM achieves strong real-world performance with substantially lower latency than imagine-then-execute variants, while removing video co-training degrades both success rate and completion time.

5 Conclusion

The demonstrated design uses video prediction to train a policy representation while avoiding its expensive explicit synthesis at deployment. It preserves high benchmark success with much lower latency than two imagination-based variants, although sequential imagination improves real-task success in the reported comparison. Larger-scale pretraining and model scaling remain future work; the results do not settle whether explicit look-ahead helps on tasks requiring search, counterfactual reasoning or unseen dynamics.

Appendix A Appendix

A.1 RoboTwin Detailed Results

Table 3 preserves all 50 RoboTwin task rows for seven methods under clean and randomized conditions. The averages conceal meaningful task differences: Fast-WAM exceeds Joint/IDM on randomized Press Stapler (97% versus 50/57%) and Place Can Basket (69% versus 23/28%), but trails them on randomized Handover Block (81% versus 91/94%) and Move Stapler Pad (64% versus 81/85%). Hanging Mug and Open Microwave remain difficult. The no-video variant even exceeds Fast-WAM on some cells, so the auxiliary objective improves the aggregate without winning every individual task.

Table 3: Per-task success rates on RoboTwin under clean and randomized evaluation settings.
Complete original PDF Table 3
Original selectable table
Task Fast-WAM (Ours) Fast-WAM-Joint Fast-WAM-IDM Fast-WAM w.o. co-train LingBot-VA π0.5\pi_{0.5} Motus
Clean Rand. Clean Rand. Clean Rand. Clean Rand. Clean Rand. Clean Rand. Clean Rand.
Adjust Bottle 100 100 98 99 94 99 98 100 90 94 100 99 89 93
Beat Block Hammer 99 97 100 98 98 98 80 92 96 98 96 93 95 88
Blocks Ranking RGB 100 100 100 100 100 99 88 86 99 98 92 85 99 97
Blocks Ranking Size 94 98 83 91 79 90 56 62 94 96 49 26 75 63
Click Alarmclock 100 100 100 100 98 100 100 98 99 100 98 89 100 100
Click Bell 100 100 100 98 100 96 100 100 100 100 99 66 100 100
Dump Bin Bigbin 97 96 95 95 93 98 92 94 89 96 92 97 95 91
Grab Roller 100 100 100 100 100 100 100 100 100 100 100 100 100 100
Handover Block 95 81 93 91 97 94 58 46 99 78 66 57 86 73
Handover Mic 99 100 100 100 98 99 100 100 94 96 98 97 78 63
Hanging Mug 58 62 71 56 66 62 28 40 40 28 18 17 38 38
Lift Pot 100 100 100 100 100 100 92 90 100 99 96 85 96 99
Move Can Pot 90 88 97 99 97 100 80 68 94 97 51 55 34 74
Move Pillbottle Pad 100 99 99 100 98 100 88 96 99 99 84 61 93 96
Move Playingcard Away 100 100 100 100 99 100 94 96 100 99 96 84 100 96
Move Stapler Pad 77 64 85 81 89 85 64 78 91 79 56 42 83 85
Open Laptop 98 100 89 92 92 92 100 98 92 94 90 96 95 91
Open Microwave 62 45 3 14 54 53 46 52 82 86 34 77 95 91
Pick Diverse Bottles 80 85 86 87 87 89 58 62 89 82 81 71 90 91
Pick Dual Bottles 100 96 98 99 100 98 80 74 100 99 93 63 96 90
Place A2B Left 95 93 96 96 97 96 84 92 97 93 87 82 88 79
Place A2B Right 93 99 95 95 94 98 88 84 97 95 87 84 91 87
Place Bread Basket 91 93 89 94 91 97 74 76 97 95 77 64 91 94
Place Bread Skillet 90 93 90 93 90 95 98 84 95 90 85 66 86 83
Place Burger Fries 96 99 100 100 97 99 94 96 97 95 94 87 98 98
Place Can Basket 71 69 50 23 37 28 72 72 81 84 62 62 81 76
Place Cans Plasticbox 99 96 98 98 98 96 98 96 100 99 94 84 98 94
Place Container Plate 96 100 99 98 100 96 94 98 99 97 99 95 98 99
Place Dual Shoes 94 88 93 89 85 87 80 74 94 89 75 75 93 87
Place Empty Cup 100 100 100 100 100 100 100 100 100 100 100 99 99 98
Place Fan 96 96 99 96 97 95 80 88 99 93 87 85 91 87
Place Mouse Pad 83 89 96 91 97 93 64 76 93 96 60 39 66 68
Place Object Basket 89 88 86 81 87 82 82 90 91 88 80 76 81 87
Place Object Scale 90 97 96 99 99 99 86 80 96 95 86 80 88 85
Place Object Stand 90 94 92 98 96 100 82 92 99 96 91 85 98 97
Place Phone Stand 97 99 100 100 99 99 90 92 97 97 81 81 87 86
Place Shoe 96 99 95 97 95 98 92 90 98 98 92 93 99 97
Press Stapler 90 97 52 50 50 57 80 80 85 82 87 83 93 98
Put Bottles Dustbin 95 90 93 95 97 92 78 88 87 91 84 79 81 79
Put Object Cabinet 94 89 95 90 93 90 88 84 85 87 80 79 88 71
Rotate QRcode 93 89 91 92 91 86 82 78 96 91 89 87 89 73
Scan Object 89 92 92 92 93 90 78 86 96 91 72 65 67 66
Shake Bottle 100 100 100 100 100 100 100 100 100 97 99 97 100 97
Shake Bottle Horizontally 100 100 100 100 100 100 100 100 100 99 99 99 100 98
Stack Blocks Three 95 97 98 97 99 95 90 90 99 98 91 76 91 95
Stack Blocks Two 100 100 100 100 100 100 100 98 100 98 97 100 100 98
Stack Bowls Three 80 81 84 86 85 83 66 82 86 83 77 71 79 87
Stack Bowls Two 92 98 97 95 94 96 90 98 94 98 95 96 98 98
Stamp Seal 90 94 96 99 99 94 60 78 96 97 79 55 93 92
Turn Switch 61 59 73 72 59 74 66 66 44 45 62 54 84 78
Average 91.88 91.78 90.84 90.32 91.16 91.34 82.76 84.80 92.90 91.50 82.74 76.76 88.66 87.02
Concept demonstration

Open image ↗