World-action models adapt pretrained video generators, and vision-language-action models build on pretrained vision-language backbones; both tie policy architecture and inference cost to large transformers designed for generation rather than control. It remains unclear how much these generation-oriented choices affect control. In this work, we introduce Latent Action JEPA (La-JEPA), a compact world-action model whose policy transformer is trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder. Following Fast-WAM, La-JEPA learns to predict future features jointly with actions during training and generates only actions at inference. Among policies without embodied pretraining, La-JEPA achieves state-of-the-art performance in clean and randomized RoboTwin 2.0 scenes and on LIBERO-Plus with correct instructions, while using far fewer parameters than the strongest competing policies. It runs substantially faster than Fast-WAM at comparable performance and outperforms the closest JEPA-based methods, LeapBot-WA and JEPA-WAM, on LIBERO-Plus and RoboTwin 2.0 without their pretrained policy backbones. On four real-robot tasks, its average success is comparable to pi0.5 and above non-embodied-pretrained baselines. Through a design study of the choices behind La-JEPA, we show that JEPA features outperform video VAE latents, most clearly out of distribution, and that training-time future prediction choices significantly impact the policy even when future prediction is skipped at inference. We correct the LIBERO-Plus instructions, as the officially released evaluator corrupts them for most variants, and re-evaluate several strong open baselines. Without a generative pretrained policy backbone, La-JEPA is better and faster than far larger policies, suggesting that robot foundation models may benefit from architectures designed for control rather than inherited from generation.
Figures & tables
Figure 1: Success against deployed parameters, including frozen encoders, on LIBERO-Plus with corrected instructions (Section 5.2 ) and on RoboTwin 2.0, where the panel reports the equally weighted mean of clean and randomized success of clean-trained policies. S, B, and L denote the La-JEPA sizes. The LIBERO-Plus panel shows the policies without embodied pretraining that we re-evaluated with corrected instructions, and the embodied-pretrained π0.5 ; Table 1 lists all scores, including those with the released instructions. ∗ : success estimated from JEPA-WAM’s 20-task results (Appendix B ).
Figure 2: La-JEPA. Left: general architecture. Right: context and future encoding with V-JEPA 2.1.
Figure 3: Gain in success (percentage points) from each design choice over its baseline: A, action-only training; B, Wan2.2 VAE latents; C, static encoding; D, no camera dropout (LIBERO: wrist-only dropout); E, no proprioception dropout; F, a plain policy without augmentation or encoder LoRA. Green: LIBERO; orange: RoboTwin 2.0; solid: in-distribution; hatched: out-of-distribution. Block A averages LIBERO-Plus gains over three training seeds. Values and evaluation budgets in Table 6 .
Params (B)
LIBERO-Plus
Method
Transformer init.
Train.
Depl.
LIBERO
Corrected
Released
Qwen-RobotManip-scratch
VLM (supervised)
4.7
4.7
98.2
—
78.3
StarVLA
VLM (supervised)
7.8
7.8
96.9
77.6
76.7
JEPA-WAM
VLM (supervised)
0.55
1.35
96.7
77.8
77.1
Fast-WAM
video generator
6.0
12.4
97.6
70.8
51.5
Faster-WAM
video generator
5.1
11.4
98.5
—
75.1
Table 1: LIBERO and LIBERO-Plus success (%) of policies without embodied pretraining (all use pretrained perception). Parameters in billions. LIBERO-Plus: episode-weighted success over all 10,030 variants with corrected or released instructions (Section 5.2 ); corrected-instruction scores exist only for the policies we re-evaluated. Best in bold, second best underlined unless the best is tied. La-JEPA values are means over three training seeds (Appendix B ). § Scored at the upstream step limits used for all policies; with its own longer limits (400 steps, 700 for LIBERO-Long), NowWAM reports 87.7% on the released instructions.
Params (B)
RoboTwin 2.0 success (%)
Method
Transformer init.
Train.
Depl.
Clean
Randomized
Mean
Qwen-RobotManip-scratch
VLM (supervised)
4.7
4.7
71.6
22.6
47.1
StarVLA
VLM (supervised)
7.8
7.8
58.1
10.6
34.4
Fast-WAM
video generator
6.0
12.4
77.8
1.9
39.9
Faster-WAM
video generator
5.1
11.4
68.6
2.2
35.4
La-JEPA S (ours)
scratch
0.11
0.75
85.5
33.9
59.7
Table 2: RoboTwin 2.0 success of clean-trained policies without embodied pretraining, in clean and randomized scenes. Parameters in billions, for the RoboTwin configurations. Mean: equally weighted mean of clean and randomized success. Best in bold, second best underlined. La-JEPA values are means over three training seeds, each evaluated on 5,000 episodes per condition (Appendix B ).
Params (B)
LIBERO
RoboTwin 2.0 (20 tasks)
Method
Transformer init.
Embodied PT
Train.
Depl.
IID
Plus
C2C
C2R
LDA-1B ‡
—
yes
—
—
90.6
45.5
—
—
LeapBot-WA
video generator
yes
2.7
10.4
97.3
71.7
—
—
JEPA-WAM
VLM (supervised)
no
0.55
1.35
96.7
77.1
79.9
36.9
La-JEPA S (ours)
scratch
no
0.10
0.74
97.9
80.0
95.8
49.9
La-JEPA B (ours)
scratch
no
0.34
0.98
98.0
82.5
95.9
49.9
Table 3: Latent world-action models. Embodied PT: pretraining on robot data beyond the target demonstrations. ‡ Reported by Yan et al. (2026b) . La-JEPA LIBERO and LIBERO-Plus values are means over three training seeds (Appendix B ).
Figure 4: RoboTwin clean versus randomized success of clean-trained policies. Blue: La-JEPA B; brown: no embodied pretraining; gray: embodied pretraining; RoboManip: Qwen-RobotManip. Points above the dashed line through La-JEPA have a higher mean of the two regimes. ∗ : JEPA-WAM estimate (Appendix B ); LDA-1B as reported by Yan et al. (2026b) .
Source
Camera viewpoint (LIBERO-Object)
Lighting (LIBERO-Spatial)
Training, corrected (ours)
pick up the alphabet soup and place it in the basket
pick up the black bowl between the plate and the ramekin and place it on the plate
Released
…in the basket view 0 0 100 2 4 initstate 0
…place it on the plate light 1
Flex- π
Pick the alphabet soup …
Pick the akita black bowl …
Table 4: Instruction received for two LIBERO-Plus variants. Bold: deviations from the training instruction.
Figure 5: Real-robot tasks: cube placement, fruit picking, Push T, and dish manipulation.
Policy
Latency (ms)
Fruit
Cube
Push T
Dish manipulation
Average
Plates
Utensil 1
Utensil 2
Position
With embodied pretraining
π0.5
71.89
20
20
11
20
19
19
18
17.25
LingBot-VA
≈4760
18
20
2
6/10
3/10
2/10
1/10
—
Without embodied pretraining
Fast-WAM
88.09
19
19
14
20
19
15
12
16.00
Table 5: Real-robot successes out of 20 and benchmark policy-call latency in ms, not control-loop timing (Appendix C ); LingBot-VA’s is rounded to 10 ms. Position: full dish success with correct utensil placement; Average: mean over the four tasks. LingBot-VA dish: 10 trials. Bold: best.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Choice
Baseline
Benchmark
IID
OOD
Evaluation
A Future prediction
Action + future prediction
Action only
LIBERO
+ 0.9
+ 2.1
IID: seed 42; OOD: 3 seeds
Action + future prediction
Action only
RoboTwin
+ 7.0
+ 3.7
5,000 /cond., 120–150k
B V-JEPA features
V-JEPA 2.1 features
Wan2.2 VAE latents
LIBERO
+ 2.0
+ 6.0
2,000 / 10,030 , 150k
V-JEPA 2.1 features
Wan2.2 VAE latents
RoboTwin
+ 15.9
+ 16.0
5,000 /cond., 150k
Appendix
Table 6: Gain in success (points) of each design choice over its baseline; comparisons hold within a block.
Seed
Selected step
Joint
Action only
Gain (pp)
Paired p
42
150k
76.13
74.47
+ 1.67
2.3×10−4
1
90k
76.13
74.77
+ 1.37
0.002
2
120k
77.15
73.82
+ 3.33
2.2×10−12
Mean
—
76.47
74.35
+ 2.12
—
Appendix
Table 7: Predictive supervision across training seeds on LIBERO-Plus (success in %; per-seed exact McNemar tests).
Training variant
Camera viewpoints
Sensor noise
Frozen encoder, stored features
Wrist-camera dropout only
11.4
42.0
Either-camera dropout
43.8
67.1
Online encoder, either-camera dropout
Image augmentation only
74.6
81.4
Encoder LoRA only
43.9
73.3
Appendix
Table 8: LIBERO-Plus success under camera-viewpoint and sensor-noise perturbations (%).
Policy size
S
B
L
Width / layers / heads
384 / 12 / 6
768 / 12 / 12
1024 / 24 / 16
LoRA blocks (LIBERO)
Last 8
Last 8
All 24
LoRA blocks (RoboTwin)
Last 8
Last 8
Last 8
Transformer / losses
Equal World/Action DiT widths and depths; MLP expansion 4; action and future loss weights 1 each
Table 10: Observation augmentation. Probabilities are conditional on not bypassing augmentation; fog is one of three corruption types.
LIBERO-Plus categories (revised)
LIBERO-Plus
LIBERO
Size
Camera
Robot
Language
Light
Backgr.
Noise
Layout
Revised
Orig.
Revised
Orig.
S
80.4
58.3
89.1
95.5
66.0
89.6
80.3
80.0 ± 0.4
75.2
97.9 ± 0.1
93.4
B
85.0
58.0
91.0
96.7
71.4
93.2
82.1
82.5 ± 0.4
78.9
98.0 ± 0.2
96.4
L
86.4
61.4
92.1
97.4
74.9
94.6
82.1
84.1 ± 0.4
82.3
98.1 ± 0.2
97.0
Appendix
Table 11: Revised versus original LIBERO policies (success in %). Revised: final 150k checkpoint, mean over three training seeds ± sample standard deviation across seeds; LIBERO-Plus categories are three-seed means. Original: one training run at 120k updates. Both use the upstream suite-specific step limits; the original LIBERO values are counted from recorded rollouts without settling steps, whereas the revised evaluations take 10 settling steps before the policy acts.
Clean
Randomized
Mean
Size
Revised
Orig.
Revised
Orig.
Revised
Orig.
S
85.5 ± 0.5
80.2
33.9 ± 2.2
39.0
59.7 ± 1.3
59.6
B
86.2 ± 0.5
78.8
35.4 ± 0.6
44.7
60.8 ± 0.5
61.8
L
86.0 ± 0.4
78.4
32.1 ± 2.8
39.1
59.0 ± 1.6
58.7
Appendix
Table 12: Revised versus original RoboTwin 2.0 policies (clean-trained; success in %, 5,000 episodes per condition). Revised: mean over three training seeds ± sample standard deviation across seeds. Original: one training run.
Source
Example instruction
Match
LIBERO training data
pick up the black bowl between the plate and the ramekin and place it on the plate
—
Released LIBERO-Plus
pick up the black bowl between the plate and the ramekin and place it on the plate light 1
0
Flex- π (BDDL annotation)
Pick the akita black bowl between the plate and the ramekin and place it on the plate
2,351
Ours (corrected)
pick up the black bowl between the plate and the ramekin and place it on the plate
8,493
Appendix
Table 13: Instruction sources for LIBERO-Plus. Example: the first LIBERO-Spatial task under the first lighting perturbation. Match: non-language variants whose instruction equals the training instruction, out of 8,493.
Policy
Embodied PT
Released
Corrected
Δ
La-JEPA S (ours)
no
80.0
86.3
+6.3
La-JEPA B (ours)
no
82.5
87.3
+4.9
La-JEPA L (ours)
no
84.1
88.0
+3.9
NowWAM
no
86.2
86.2
+0.0
Cosmos Policy §
no
80.8
84.9
+4.1
ImageWAM
no
79.7
80.4
+0.7
Appendix
Table 14: LIBERO-Plus success (%) with the released and the corrected instructions, all evaluated in our harness on all 10,030 variants (MuJoCo 3.3.2, seed 0, first initial state, 10 settling steps, upstream per-suite step limits). Δ : corrected minus released. Best in bold, second best underlined. La-JEPA values are means over three training seeds. § One variant cannot be instantiated by LIBERO-Plus, so n=10,029 . ¶ Run-to-run variation of about ±0.5 points. π -JEPA-WAM: π0.5 trained with the auxiliary JEPA objective of JEPA-WAM ( Lin et al., 2026 ) .
Policy
Steps
Optimized
Original-kernel CUDA graphs
Eager
La-JEPA S
10
17.78
34.20
99.96
La-JEPA B
10
25.16
42.35
98.21
La-JEPA L
10
59.72
88.87
167.93
Fast-WAM
10
88.09
OOM
193.44
Faster-WAM
10
50.96
OOM
62.70
π0.5
10
71.89
—
—
Appendix
Table 15: Policy-call latency (ms): batch size one, median of 100 calls after 20 warm-ups on one consumer GPU. OOM: out of memory; dashes: not measured. The timed L uses last-eight-block rather than full-depth LoRA, which does not change latency.
Method
Trained components
Frozen components at inference
Trainable
Deployed
Qwen-RobotManip-scratch
Qwen3.5-4B, action head
—
4.7 e
4.7 e
StarVLA
Qwen3-VL-4B (4.44), cross-attention DiT action head (3.36)
Figure 6: Companion to Figure 1 : in-distribution (top) and out-of-distribution (bottom) success against trainable parameters on a logarithmic axis, for LIBERO (left) and clean-trained RoboTwin 2.0 (right). Trainable counts every weight updated on the target demonstrations, adapters included. La-JEPA points are blue and connected in size order; purple denotes VLM-initialized and brown generator-initialized action transformers, while other scratch policies use muted blue. SmolVLA ( Shukor et al., 2025 ) is shown at its three sizes joined by a dotted line. Unlike Figure 1 , LIBERO-Plus uses the released instructions, for which scores of all baselines are available.
With embodied pretraining
Without embodied pretraining
Task
π0.5
LingBot-VA
Fast-WAM
Faster-WAM
La-JEPA
Fruit
24.4 ± 2.7
175.3 ± 41.0
36.7 ± 9.0
35.1 ± 9.5
21.4 ± 0.8
Cube
10.8 ± 0.5
117.8 ± 50.9
21.9 ± 7.6
18.2 ± 4.6
11.5 ± 3.2
Push T
22.8 ± 14.7
112.5 ± 9.1
24.6 ± 6.6
23.8 ± 12.1
21.1 ± 4.1
Dish
39.0 ± 3.4
269.0 †
85.0 ± 15.3
72.0 ± 8.8
58.1 ± 8.7
Appendix
Table 17: Real-robot completion time over successful trials (s; standard deviation in small type). Bold: fastest; underlined: second. † Single successful trial.
S
B
L
Task
Clean
Rand.
Clean
Rand.
Clean
Rand.
Adjust bottle
100.0
21.0
100.0
18.7
99.7
15.7
Beat block hammer
91.3
17.7
94.7
20.7
94.0
21.0
Blocks ranking rgb
95.7
35.7
95.0
43.7
97.0
43.3
Blocks ranking size
67.3
29.3
69.3
34.0
66.7
29.7
Click alarmclock
99.7
47.7
100.0
68.0
99.7
64.3
Appendix
Table 18: Per-task RoboTwin 2.0 success of La-JEPA S, B, and L (%), mean over three training seeds with 100 episodes per task, condition, and seed.
Model
Evaluation setting
Value
Success (%)
S
Reference
Default
94.30
S
Denoising steps
3
94.50
L
Reference
Default
97.40
B
Reference
Default
97.60
L
Execution chunk
5
97.00
L
Execution chunk
8
96.30
Appendix
Table 19: Evaluation-setting sweeps at fixed checkpoints (success in %). LIBERO values use the 700-step budget in every suite (Appendix B ); B’s LIBERO reference is a re-evaluation of its default setting.
Future pooling
IID
OOD
2× head, 4× wrists
61.24
15.68
4× head, 8× wrists
61.96
18.56
Global spatial pooling
61.14
15.04
Appendix
Table 20: Future-only pooling on RoboTwin (success in %).
Training objective
Clean success (%)
Effective rank
Action readout R2
Action + future prediction
66.5
137.70
0.4315
Action only
62.8
48.41
0.7937
Appendix
Table 21: Representation diagnostics with and without future prediction on RoboTwin.
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
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
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