The vision-language-action (VLA) and world-action model (WAM) communities debate whether control policies should inherit pretrained knowledge from vision-language models or from visual generators. The debate, however, bundles three separable choices: backbone initialization, the objective applied to future observations, and whether the machinery for that future loss is kept at inference. A gap between a VLA and a WAM therefore cannot simply be attributed to any one of them. We show that decoupling these choices yields a common framework for both families, which we term Just Action Model (JAM): a pretrained visual backbone with an action expert, trained with an action loss and a foresight loss. The foresight loss trains current-observation features to predict frozen encodings of the recorded future through a prediction head discarded after training, so no visual generation is involved in either training or inference, and the backbone can be chosen freely. We first compare objectives within fixed policy architectures, then evaluate the foresight recipe across backbone families. The resulting JAM policies outperform their matched FastWAM and ImageWAM counterparts on LIBERO and LIBERO-Plus, and achieve similar or slightly higher success rates on RoboTwin. Compared with FastWAM, JAM improves the success rates from 97.0%/55.3%/91.2% to 98.2%/64.5%/91.4%, respectively. With cached foresight targets, JAM reduces estimated online training FLOPs per sample per update by 73% relative to FastWAM. Our results show that future supervision improves control without visual generation or additional inference machinery, and strong foresight performance is attainable with either generator-initialized or vision-language backbones. Project Page: https://y0uroy.github.io/JAM
Figures & tables
Figure 1: The three choices that separate VLAs, WAMs, and JAM: backbone initialization, the objective on recorded future observations, and the modules kept at inference. (a) The VLA reference uses no future objective, only the action loss Lact . (b) WAMs add visual generation loss LG . The dashed loop shows generated future observations entering the action expert in some WAMs. (c) JAM uses foresight loss LF across pretraining families, discards its supervision branch, and retains the corresponding action path.
Figure 2: JAM training (left) and deployment (right), with the notation of Section 3.2 . The input and text encoders stand for architecture-specific interfaces. Recorded futures f1:T enter only the future anchor branch inside the dashed box; deployment discards this branch and the prediction head, leaving the backbone and action expert.
Figure 3: Matched LIBERO-Long rollouts for ImageWAM and JAM, both on Ovis-U1.
Policy
LG
LF
Future anchor
Readout offsets
LIBERO
LIBERO-Plus
FastWAM
×
×
–
–
95.25
45.55
✓
×
–
–
97.00
55.31
JAM (Wan)
×
✓
Pretrained Wan
[0]
95.80
44.62
×
✓
[1,2]
97.45
55.58
×
✓
[0,1,2]
97.40
54.18
×
✓
Random Wan
[1,2]
97.65
58.63
Table 1: Generation versus foresight supervision with the policy architecture fixed. LG is the generative loss and LF the foresight loss. The target rows also vary the future anchor. Readout offsets are the set K of target positions, with 0 the current frame. The total foresight weight λF is fixed at 0.6 for Wan and 0.3 for Ovis-U1 and is divided equally among the readout offsets.
Figure 4: Decoded future change for put both moka pots on the stove , for ImageWAM and the two JAM (Ovis-U1) policies of Table 1 with readout offsets [0] and [1] . Two MLP readouts reconstruct the current frame and the frame 16 steps ahead from the same frozen features, and their difference is compared with the recorded change. Red marks pixels that get brighter and blue marks pixels that get darker, on a common scale. In this example, the features of JAM trained with offset [1] make the future-frame change more readily decodable than the features of ImageWAM or JAM trained with offset [0] .
Table 6
Figure 5: Action-to-visual attention on one LIBERO-Long observation for put the white mug on the left plate and put the yellow and white mug on the right plate , in the agent and wrist views. (a) ImageWAM (Ovis-U1) and JAM share the deployed policy architecture but use different training objectives. The difference map compares normalized display intensities, with red indicating higher values for JAM and blue for ImageWAM. (b) JAM on a self-supervised and a vision-language backbone. Bars show attention shares over text (orange), image (blue), and action keys (green), averaged over 100 reference frames.
Table 8
Method
P.T.
Camera
Robot
Language
Light
Background
Noise
Layout
AVG
VLM-based policies
UniVLA ( Bu et al., 2025 )
✓
1.8
46.2
69.6
69.0
81.0
21.2
31.9
42.9
OpenVLA-OFT ( Kim et al., 2025a )
✓
56.4
31.9
79.5
88.7
93.3
75.8
74.2
69.6
π0 ( Black et al., 2025b )
✓
13.8
6.0
58.8
85.0
81.4
79.0
68.9
53.6
π0 -Fast ( Pertsch et al., 2025 )
✓
65.1
21.6
61.0
73.2
73.2
74.4
68.8
61.6
WorldVLA ( Cen et al., 2025 )
×
0.1
27.9
41.6
43.7
17.1
10.9
38.0
25.0
Table 6: LIBERO-Plus success (%) by perturbation category, with the task-weighted average. Published rows retain their original evaluation protocols; † marks our evaluation on a fixed 1,506-task subset.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Future anchor
Training targets
Action inference
Frames
Layer
Readout
Context
Wan
0,4,…,32
18/30
[1,2]
Wan + UMT5
Wan VAE
Yak18
0,16
11/18
[1]
Yak18 + Ovis
FLUX.2
15/25
FLUX.2 + Qwen3-4B
Qwen3.5-2B
0,16
13/24
[1]
Causal image–text prefix
Appendix
Table 7: Training targets and action contexts across backbones. Frame offsets are measured in environment steps relative to the current observation; Layer reports the supervised policy block / total depth.
Policy
Input D
Target Dg
Trunk d
Layers
Heads
FFN
Readouts
JAM (Wan)
3072
3072
1536
2
24
6144
2
JAM (Ovis-U1)
1536
1536
768
2
12
3072
1
JAM (FLUX.2)
3072
3072
1536
2
24
6144
1
JAM (Qwen3.5-2B)
2048
2048
1024
2
16
4096
1
JAM (DINOv2-B)
768
768
384
2
6
1536
1
JAM (DINOv2-L)
1024
1024
512
2
8
2048
1
Appendix
Table 8: Default prediction-head dimensions. Layers, Heads, and FFN describe the shared transformer trunk; Readouts denotes the number of supervised offsets.
Training Params. (M)
TFLOPs per sample per update
Measured GPU time
Policy
Targets
Head
Anchor
Visual
Action
Head
Target
VAE
Total
Reduction
s / step
Reduction
FastWAM
–
–
–
8.93
0.05
–
–
6.54
15.52
–
0.789
–
JAM (Wan)
online
71
0
3.17
0.05
0.03
0
6.54
9.80
36.9%
0.532
32.6%
JAM (Wan)
cached
71
0
3.17
0.05
0.03
0
0.93
4.19
73.0%
0.258
67.3%
ImageWAM (Ovis-U1)
–
–
–
4.91
0.07
–
–
0.83
5.81
–
0.473
–
JAM (Ovis-U1)
online
16.5
1,127
3.13
0.07
0.05
0.80
0.83
4.88
15.9%
0.403
14.8%
Appendix
Table 9: Training-only parameters, estimated online FLOPs per sample per update, and measured GPU time over the profiled operations.
Policy
Future anchor
Spatial
Object
Goal
Long
AVG
FastWAM
–
97.4
99.6
96.6
94.4
97.00
FastWAM action-only
–
95.0
99.6
94.4
92.0
95.25
JAM (Wan)
Pretrained Wan
98.2
100.0
97.6
94.0
97.45
Online Wan EMA
95.0
99.6
96.4
94.6
96.40
Appendix
Table 10: Frozen pretrained and EMA Wan future anchors on LIBERO. Success rates are in percent.
Policy
Depth
Spatial
Object
Goal
Long
AVG
FastWAM
–
97.4
99.6
96.6
94.4
97.00
FastWAM action-only
–
95.0
99.6
94.4
92.0
95.25
JAM (Wan)
14
98.0
99.8
96.4
94.2
97.10
16
99.2
99.8
96.2
94.0
97.30
18
98.2
100.0
97.6
94.0
97.45
20
96.8
99.2
95.4
90.4
95.45
Appendix
Table 11: LIBERO success in percent by Wan block. Depth indexes the 30 visual blocks.
Policy
λF
Spatial
Object
Goal
Long
AVG
FastWAM
–
97.4
99.6
96.6
94.4
97.00
FastWAM action-only
–
95.0
99.6
94.4
92.0
95.25
JAM (Wan)
0.4
96.8
99.4
96.6
90.0
95.70
0.5
95.6
99.8
95.2
97.0
96.90
0.6
98.2
100.0
97.6
94.0
97.45
0.7
97.4
99.8
96.0
93.8
96.75
Appendix
Table 12: LIBERO success rates in percent across foresight loss weights λF for Wan.
Policy
Head
Width
FFN
Params (M)
Spatial
Object
Goal
Long
AVG
FastWAM
–
–
–
–
97.4
99.6
96.6
94.4
97.00
FastWAM action-only
–
–
–
–
95.0
99.6
94.4
92.0
95.25
JAM (Wan)
Transformer
3072
4096
145
97.4
100.0
96.2
91.6
96.30
1536
6144
71
98.2
100.0
97.6
94.0
97.45
768
6144
31
96.4
100.0
97.0
89.6
95.75
MLP
3072
–
38
95.8
100.0
94.2
91.4
95.35
Appendix
Table 13: Prediction-head capacity on LIBERO. Heads use 3,072-D inputs and outputs, two-layer trunks, and two readouts. Transformers have 24 attention heads. Their input projections are learned except at width 3,072, which uses the identity.
Method
Camera
Robot
Language
Light
Background
Noise
Layout
AVG
Qwen3.5-2B action-only
80.00
68.02
72.54
92.55
54.84
93.02
75.45
77.03
JAM (Qwen3.5-2B)
80.83
61.42
71.72
93.79
81.72
94.19
84.55
81.14
DINOv2-L action-only
37.92
32.99
86.48
86.96
85.48
78.29
72.73
68.26
JAM (DINOv2-L)
55.42
39.09
80.74
85.71
77.42
81.40
73.18
70.39
Appendix
Table 14: Action-only controls on LIBERO-Plus. Success rates are in percent; AVG is micro-averaged. Bold marks the higher score within each backbone pair.
Policy
Actions
Params
Camera
Robot
Language
Light
Background
Noise
Layout
AVG
JAM (Qwen3.5-2B)
Expert
3.60B
80.8
61.4
71.7
93.8
81.7
94.2
84.5
81.14
JAM (Qwen3.5-2B)
Shared
2.22B
80.4
64.0
68.9
92.5
76.9
96.1
78.2
79.61
JAM (Qwen3.5-4B)
Shared
4.55B
80.4
67.5
74.6
91.9
76.3
95.3
77.7
80.68
Appendix
Table 15: Separate action expert versus shared backbone weights for Qwen3.5 on LIBERO-Plus. Success rates are in percent; AVG is task-weighted and Params counts deployed parameters.
Policy
Updates
LIBERO
LIBERO-Plus
FastWAM
10 k
95.90
55.25
20 k
97.00
55.31
JAM (Wan)
10 k
97.45
48.47
20 k
97.45
55.58
Appendix
Table 16: Success rates in percent for FastWAM and JAM with Wan at two training budgets. Each run has its own full learning-rate schedule, so the 10 k rows are independently trained endpoints.
Backbone
Anchor
Spatial
Object
Goal
Long
AVG
Pretrained
Pretrained
98.2
100.0
97.6
94.0
97.45
Pretrained
Random
98.0
99.8
97.0
95.8
97.65
Random
Random
85.6
98.0
88.2
65.8
84.40
Random
Pretrained
76.0
95.0
70.4
43.2
71.15
Appendix
Table 17: LIBERO success rates in percent for pretrained and random initialization of the Wan backbone and future anchor.
Method
P.T.
Spatial
Object
Goal
Long
AVG
VLM-based policies
OpenVLA ( Kim et al., 2025b )
✓
84.7
88.4
79.2
53.7
76.5
GR00T N1 ( NVIDIA et al., 2025 )
✓
94.4
97.6
93.0
90.6
93.9
π0 ( Black et al., 2025b )
✓
96.8
98.8
95.8
85.2
94.1
π0.5 ( Black et al., 2025a )
✓
98.8
98.2
98.0
92.4
96.9
JAM (Qwen3.5-2B, Qwen3.5 targets) †
×
97.8
99.6
97.2
91.6
96.55
Appendix
Table 18: LIBERO success rates in percent by suite. AVG is the four-suite mean. Published rows follow FastWAM ( Yuan et al., 2026 ) , except GR00T N1, reported by LingBot-VA ( Li et al., 2026b ) ; they retain their source protocols.
Figure 6: Paired LIBERO-Plus outcomes on 1,506 shared tasks. Left bars count tasks solved only by the baseline; right bars count tasks solved only by JAM. Panel a uses Ovis-U1, and panel b uses Wan with VAE-direct [1,2] targets.
Method
P.T.
Clean
Rand.
AVG
VLM-based policies
π0 ( Black et al., 2025b )
✓
65.92
58.40
62.16
π0.5 ( Black et al., 2025a )
✓
82.74
76.76
79.75
JAM (Qwen3.5-2B) †
×
89.70
89.86
89.78
Video generator-based policies
Motus ( Bi et al., 2026 )
✓
88.66
87.02
87.84
Appendix
Table 19: RoboTwin 2.0 success rates in percent. AVG is the mean of the Clean and Randomized settings. Published Clean and Randomized values are from FastWAM ( Yuan et al., 2026 ) , Table 1; rows marked † are evaluated by us.
Figure 9: Mean future-frame PSNR.
Policy family
Benchmark
Δ
95% task-bootstrap CI
Baseline only / JAM only
McNemar p
FastWAM
LIBERO
+1.15
[+0.20,+2.10]
30/53
0.015
ImageWAM
LIBERO
+3.85
[+1.00,+7.40]
42/119
1.01×10−9
FastWAM
LIBERO-Plus
+9.23
[+6.84,+11.69]
111/250
1.91×10−13
ImageWAM
LIBERO-Plus
+14.67
[+12.15,+17.13]
89/310
9.69×10−30
Appendix
Table 20: Fixed-checkpoint comparisons with generative baselines, using VAE-direct [1,2] targets for Wan and pretrained Yak [1] for Ovis-U1. Δ is the gain in percentage points; counts are successes unique to each policy.
Policy
Run
Spatial
Object
Goal
Long
AVG
LIBERO-Plus
JAM (Wan)
1
98.80
100.00
98.00
95.80
98.15
64.54
2
98.60
99.80
98.20
95.40
98.00
64.28
3
99.00
100.00
97.80
96.20
98.25
65.01
Mean ± SD
98.80±0.20
99.93±0.12
98.00±0.20
95.80±0.40
98.13±0.13
64.61±0.37
JAM (Ovis-U1)
1
98.60
99.60
98.40
90.40
96.75
79.95
2
98.80
99.60
98.40
90.80
96.90
80.35
Appendix
Table 21: Success rates in percent across three independent training runs. Summaries give mean ± sample standard deviation, computed before rounding.
Robust robot control benefits from explicitly modeling state transitions, but video-generation world action models (WAMs) introduce substantial deployment cost. Existing latent WAMs avoid explicit future generation, but often compress predictive representations or separate predictive modeling from the representations used for action generation. We introduce JEPA-WAM, a latent WAM built in a pretrained V-JEPA space, which couples latent transition prediction with continuous action generation through a shared predictor. JEPA-WAM predicts a spatially structured joint current-future target that captures task-shared visual temporal structure between current and future observations, while preserving dense patch-level correspondence. Through the shared predictor, transition supervision directly shapes the backbone, from which dedicated representations are extracted for action prediction. The same design can also be instantiated in pretrained VLA policies while preserving their original perception and action pathways. On LIBERO-Plus, JEPA-WAM achieves 79.2%, the best result without large-scale robot-policy pretraining, while its pretrained π0.5 instantiation reaches 86.3%, achieving the best overall performance. Experiments on RoboTwin 2.0 and real-world bimanual manipulation further demonstrate strong generalization under visual and spatial shifts.
Yihan Lin, Jiawei He, Shifeng Bao +6
School of Information, Renmin University of China, Beijing, China · XYZ Embodied AI, Beijing, China · Key Laboratory of Data Engineering and Knowledge Engineering, Beijing, China +3
World-action models (WAMs) couple future visual-state prediction with action generation. By adapting video generators or image-editing models pretrained at scale, a prominent line of recent WAMs inherits both predictive knowledge and the models in which it was learned. We ask whether a predictive visual latent space induced by large-scale predictive pretraining can instead provide a sufficient foundation for effective WAM learning without inheriting a complete pretrained visual generative model. To answer this question, we introduce V-JEPA Policy, a simple framework that builds a WAM on the latent space of a frozen V-JEPA 2.1 encoder. An instruction-conditioned future-latent predictor and a flow-matching action expert are jointly learned from scratch in a single downstream stage, with the predictor's future-informed context key--value states conditioning action generation. With 0.9B total parameters, of which 0.6B are trainable, V-JEPA Policy achieves competitive performance with representative WAM and vision-language-action baselines across LIBERO, LIBERO-Plus, and RoboCasa-GR1. Comparing visual foundations under the same downstream framework and training budget identifies V-JEPA latents as more effective than the discriminative, reconstructive, and video-understanding-oriented alternatives, particularly under distribution shifts. Beyond task-specific learning, pretraining the predictor on DROID video--instruction pairs without action labels and adapting it into a WAM yields substantial gains in downstream control and out-of-distribution generalization. Together, these findings establish predictive visual latents as a foundation for effective WAM learning from task-specific demonstrations and for transferring future-modeling knowledge acquired from broader in-the-wild videos. Project page: https://breez3young.github.io/vjepa-policy/
Yang Zhang, Jiangyuan Zhao, Chenyou Fan +4
Tsinghua University · Shanghai Jiao Tong University · Fudan University +4
Vision-Language-Action models (VLAs) leverage large-scale vision-language pretraining for semantic robot control, but often lack explicit foresight into how robot actions change the scene. World-Action Models (WAMs) address this limitation by conditioning policies on predicted futures, yet existing approaches typically rely on computationally expensive video generation with substantial pixel-level redundancy. We present LaWAM, a Latent World Action Model that exposes predictive dynamics to robot policies through compact latent visual subgoals instead of reconstructed future video. At the core of LaWAM is a latent-action-conditioned Latent World Model (LaWM). We obtain LaWM by training a latent action model in the latent space of a pretrained vision foundation model and repurposing its forward decoder to predict future observation features for scene evolution. LaWAM then conditions action generation on these predicted latent visual subgoals to enable dynamics-aware robot control. LaWAM achieves state-of-the-art or competitive success rates (SRs) across LIBERO (98.6% SR), RoboTwin (91.22% SR), and real-world manipulation tasks while retaining low-latency inference. LaWAM runs in 187 ms per action-chunk prediction and achieves up to 24x lower wall-clock latency than pixel-space WAMs.
Jialei Chen, Kai Wang, Kang Chen +9
Jilin University · Zhongguancun Academy · Nankai University +4