Robot action planning in the real world is challenging as it requires not only understanding the current state of the environment but also predicting how it will evolve in response to actions. Vision-language-action (VLA), which repurpose large-scale vision-language models for robot action generation using action experts, have achieved notable success across a variety of robotic tasks. Nevertheless, their performance remains constrained by the scope of their training data, exhibiting limited generalization to unseen scenarios and vulnerability to diverse contextual perturbations. More recently, world models have been revisited as an alternative to VLAs. These models, referred to as world action models (WAMs), are built upon world models that are trained on large corpora of video data to predict future states. With minor adaptations, their latent representation can be decoded into robot actions. It has been suggested that their explicit dynamic prediction capacity, combined with spatiotemporal priors acquired from web-scale video pretraining, enables WAMs to generalize more effectively than VLAs. In this paper, we conduct a comparative study of prominent state-of-the-art VLA policies and recently released WAMs. We evaluate their performance on the LIBERO-Plus and RoboTwin 2.0-Plus benchmarks under various visual and language perturbations. Our results show that WAMs achieve strong robustness, with LingBot-VA reaching 74.2% success rate on RoboTwin 2.0-Plus and Cosmos-Policy achieving 82.2% on LIBERO-Plus. While VLAs such as π0.5 can achieve comparable robustness on certain tasks, they typically require extensive training with diverse robotic datasets and varied learning objectives. The evaluation code for the RoboTwin2.0-Plus benchmark is available at: https://robot-robustness.github.io/RoboTwin2.0-Plus/.
Figures & tables
Model
#Params
Backbone
MOT
Pretrain Free
Causal Pred.
AR Gen.
VPP ( Hu et al., 2025 )
1.5B
Stable Video Diffusion (SVD)
✗
✗
✓
✗
GE-Act ( Liao et al., 2025 )
2.2B
LTX-Video-2B
✓
✗
✗
✓
Cosmos-Policy ( Kim et al., 2026 )
2B
Cosmos-Predict2-2B
✗
✓
✗
✗
LingBot-VA ( Li et al., 2026b )
5.3B
Wan2.2-5B
✓?
✗
✓
✓
DreamZero ( Ye et al., 2026b )
14B
Wan2.1-14B
✗
✗
✗
✓
GigaWorld-Policy ( Ye et al., 2026a )
>5B
Wan2.2-5B
✗
✗
✓
✗
Table 1: Summary of world action models (WAMs) . MOT : mixture-of-transformers, indicating that separate transformer backbones are used for video and action streams, with cross-modal interactions facilitated via attention at each layer. Note that the released LingBot-VA model adopts a unified transformer for both modalities, instead of the mixture-of-transformer architecture described in the paper. Pretrain Free : the method does not require task-agnostic embodied pre-training. Causal Pred. : action prediction is conditioned on the generated visual state, or vice versa. AR Gen. : auto-regressive generation.
Model
Semantic Understanding & Spatial Grounding
General Video
Robot Data
Task-specific Data
VLAs
π0
-
-
Cross-embodiment (>10k h)
Trajectories (5–100 h)
OpenVLA-OFT
-
-
Cross-embodiment (970k)
Trajectories (20–300)
X-VLA
-
-
Cross-embodiment (288k)
Trajectories (50)
π0.5
High-level planning
-
Mobile manipulation (400 h)
Trajectories (1–20 h)
Web data-VQA
Multi-env. tabletop
Table 2: Training data of VLAs and WAMs across different training stages. Training stages : Embodied Pre-training , Embodied Post-training , Task-specific Finetuning . PT : filtered data for embodied post-training.
Figure 1: Examples of perturbations on RoboTwin 2.0-Plus tasks. Definition of notations (N1, L1, etc.) can be found in Appendix A
Model
Original
Camera
Robot
Lang.
Light
BG
Noise
Layout
Total
VLAs
π0.5
78.4
45.6
27.6
74.4
49.6
71.7
64.9
56.8
58.6
X-VLA
65.6
23.2
65.2
64.4
63.1
58.6
49.7
34.8
53.1
VLA + WM
MOTUS
87.0
21.6
85.0
83.2
84.6
84.4
43.1
82.8
71.5
WAMs
Table 3: RoboTwin 2.0-Plus evaluation results . The Original column reports the success rate of each method evaluated in the original RoboTwin 2.0 Easy setting. For a fair comparison, all methods use a single unified model across all tasks. The best result in each column is highlighted in bold , while second-best result is underlined . Lang. : Language. BG : Background.
Figure 2: Case studies on RoboTwin 2.0-Plus. We present three representative cases comparing π0.5 and LingBot-VA. Keyframes are shown sequentially from left to right. (a) Task: beat block with hammer. Perturbation: noise (N3). Result: π0.5 collides with the hammer and fails to complete the task. (b) Task: handover block. Perturbation: layout. Result: π0.5 collides with the red block during approach, leading to failure. (c) Task: rank RGB blocks. Perturbation: lighting (L1–L4 mix). Result: π0.5 fails to grasp the first red block due to misalignment and does not recover. LingBot-VA success in all three cases shown.
Model
Original
Camera
Robot
Lang.
Light
BG
Noise
Layout
Total
VLAs
π0
94.2
13.8
6.0
58.8
85.0
81.4
79.0
68.9
53.6
π0(rerun)
91.3
61.0
40.8
63.5
89.3
84.1
80.1
76.4
69.4
π0 -FAST
85.5
65.1
21.6
61.0
73.2
73.2
74.4
68.8
61.6
π0.5
96.9
75.4
77.5
85.6
96.9
94.6
89.7
85.7
85.7
Openvla-OFT_m
97.6
55.6
21.7
81.0
92.7
91.0
78.6
68.7
67.9
Table 4: LIBERO-Plus evaluation results . The Original column denotes results on the standard LIBERO benchmark. The results of π0 , π0-FAST , Openvla-OFT-m, UniVLA, RIPT-VLA are sourced from Fei et al. (2025) . The results reported for HoloBrain0-GD ( Lin et al., 2026 ) and ABot-M0 ( Yang et al., 2026 ) are taken from their respective papers. The remaining models are evaluated using their official repositories and publicly released checkpoints. Specifically, π0(rerun) and π0.5 are evaluated using their JAX implementations. The best result in each column is highlighted in bold , while second-best result is underlined . Lang. : Language. BG : Background.
Figure 3: Illustrations of future images predicted by Cosmos-policy. Ground-truth and predicted images under three types of perturbations (noise, lighting, and background variations) in LIBERO-Plus are shown. GT: the ground truth. Pred.: the predictions generated by Cosmos-policy.
π0.5
X-VLA
Fast-WAM ∗
GE-Act
Cosmos-Policy
LingBot-VA(RW)
MOTUS
LingBot-VA(RT)
Action chunk size
50
30
32
36
16
32
16
32
Inference t (wall clock)
63 ms
195 ms
190 ms
300 ms
390 ms
480 ms
1175 ms
5230 ms
Inference t (w.r.t. π0.5 )
1.0
3.1 ×
3.0 ×
4.8 ×
6.2 ×
7.6 ×
18.6 ×
83.0 ×
Table 5: Runtime Speed Comparison. The inference times of all models are evaluated using the same device. LingBot-VA(RW): real-world setting with three denoising steps for state and five for action. LingBot-VA(RT): RoboTwin 2.0 setting with 25 denoising steps for state and 50 for action.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Dimension
Code
Sub-dimension
Implementation Details
Sensor Noise
N1
Motion blur
Gaussian kernel ( r∈[3,15] , σ∈[1,8] ), rotated by random angle ∈[−30∘,30∘]
Per-channel RGB tint ∈[0.0,3.5] applied to all directional and point lights (always on)
Appendix
Table 6: RoboTwin 2.0-Plus perturbation taxonomy. All 21 sub-dimensions across 7 LIBERO-Plus dimensions, with implementation-level parameter details.
#
Config
Perturbation
Active Sub-dims
0
demo_clean
None (clean baseline)
–
1
demo_vision_noise
Sensor Noise
N1–N5 (cycled)
2
demo_light
Lighting
L1 (always) + L2/L3/L4 (stochastic)
3
demo_camera
Camera Viewpoints
C1 + C3 (C2 disabled by default)
4
demo_robot_state
Robot Initial State
Joint noise + gripper extremes
5
demo_background_plus
Background
B1 (texture + color tint) + B2
Appendix
Table 7: RoboTwin-Plus evaluation configs. The 8 required configs for a complete robustness evaluation.
Aspect
LIBERO-Plus
RoboTwin 2.0-Plus
Simulator
MuJoCo (robosuite)
SAPIEN (ManiSkill3)
Robot
Franka Panda (7-DoF)
Aloha-AgileX (14-DoF)
Arms
Single arm
Dual arm (bimanual)
Cameras
2 (third-person + wrist)
3 (head + 2 wrist)
Image resolution
256×256
320×240
Native action space
7-dim delta EEF
14-dim joint positions
Appendix
Table 8: Benchmark comparison between LIBERO-Plus and RoboTwin 2.0-Plus.
Model
Action Repr.
Dim
Chunk
Freq
VLA Models
Pi0
Delta EEF
7
50
50 Hz
Pi0-FAST
Delta EEF (FAST tokenized)
7
50
15–50 Hz
Pi0.5
Absolute EEF / joints
18–19
50
50 Hz
OpenVLA-OFT
Delta EEF
7
8
3–10 Hz
UniVLA
Delta EEF (DCT tokenized)
7
10
–
Appendix
Table 9: Model setup comparison. Action representation, dimensionality, chunk size, and control frequency for each evaluated model. Entries marked with “–” indicate values inherited from the base model or not explicitly specified. Unless noted, models on LIBERO follow the native axis-angle rotation convention.
Category
Model
Camera
Robot
Lang.
Light
BG
Noise
Layout
Total
Spatial
π0
70.7
49.1
67.9
92.8
95.0
87.7
94.0
78.6
π0.5
76.9
85.4
90.3
97.6
95.3
91.7
97.7
90.3
X-VLA
24.5
91.4
82.1
100.0
100.0
63.5
93.8
77.7
VLA-JEPA
74.2
71.7
96.1
97.9
99.6
77.8
95.3
86.9
GE-Act
72.3
75.7
90.8
100.0
89.1
97.7
89.6
87.5
Cosmos-policy
81.1
59.7
88.2
89.0
91.1
98.0
99.2
86.6
Appendix
Table 10: Detailed LIBERO-Plus results for each task suite. Lang. : Language. BG : Background. The Fast-WAM LIBERO checkpoint was trained on clean demonstrations only, whereas its RoboTwin checkpoint was trained on clean + domain-randomized demonstrations (cf. RQ 2 ).
Vision-language-action (VLA) policies excel at generalist robotic manipulation but generate actions without explicitly modeling their environmental consequences. World action models (WAMs) incorporate future prediction, yet most rely on costly video generation, while latent alternatives either use predicted futures only as auxiliary targets or infer latent actions from suboptimal visual transitions through costly large-scale video pretraining. We introduce WAVE-0 (World-model-Augmented VLA Execution), a framework that bridges VLAs and WAMs through a lightweight latent world model. By connecting the semantic understanding of pretrained VLAs with explicit future prediction, WAVE-0 allows anticipated action consequences to directly guide action generation. Built on a pretrained VLA, its 257M world model translates VLM-generated latent action plans into predicted future representations that condition the action expert, unifying latent planning, future prediction, and action generation. The world model is trained directly on ground-truth robot actions with a counterfactual ranking objective, requiring neither pixel reconstruction nor large-scale video pretraining. WAVE-0 achieves success rates of 91.47% on RoboTwin 2.0 and 98.7% on LIBERO, along with 88.0%, 61.4%, and 75.0% on LIBERO-Plus, LIBERO-Pro, and LIBERO-Para, respectively. On real-world tasks, it improves the success rate of π0.5 from 38.3% to 73.3%. These results demonstrate strong manipulation performance and generalization across diverse distribution shifts.
Haomin Wang, Anqiong Ji, Ke Wang +4
AutoLab, School of Artificial Intelligence, Shanghai Jiao Tong University · Baidu AI Cloud · Awaken Robotics +1
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
Vision-Language-Action (VLA) models map observations to actions with no objective that accounts for how the world responds, so their robustness is bounded primarily by data coverage. World models carry precisely that missing objective and are better grounded for it, yet rolling the future forward costs seconds per decision and rules them out of the control loop. We show the two can be separated. What a world model knows about physical scenes lives in its internal features; generating the future is merely the objective that produced them, so the grounding can be inherited while the generative machinery is left behind. We add one feature-alignment term to ordinary VLA training: a frozen world model is run over the training frames once and cached, and the student learns to agree with that cache. No teacher is loaded during training, the projector is discarded after it, and the deployed policy is identical to the undistilled baseline, running in 32ms and 1.86GB on a consumer RTX5090, so every gain is attributable to the representation rather than to added capacity or test-time compute. A 0.8B student reaches 97.9% on LIBERO, improves from 48.2% to 50.5% on RoboCasa-GR1 humanoid manipulation, and the same objective carries over to real hardware, on both a single-arm and a bimanual platform. The gain survives changes of student scale, backbone, alignment layer, and teacher, indicating a broad representational prior rather than a fragile alignment between two particular networks. Project page: https://thaw-vla.trung-dt.com/.
Trung Dao, Sankalp Yamsani, Jaden Park +2
University of Wisconsin–Madison, USA. · University of Illinois Urbana-Champaign, USA.