WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction

• VLA policies map observations to actions without explicitly modeling environmental consequences, while video-based WAMs incur substantial computation predicting action-irrelevant visual detail.Latent approaches either leave predicted futures loosely connected to action generation or infer latent actions from visual transitions, with substantial pretraining costs in LaWAM.Source: §1 · pp. 1–2
• Its two-stage strategy trains the world model on action-labeled trajectories with counterfactual ranking, then integrates the frozen model into pretrained π0.5.A lightweight resampler compresses predicted futures into conditioning tokens while the action expert retains the original VLM context.Source: §1 · p. 2
• WAVE-0 reaches 91.47% on RoboTwin 2.0 and 98.7% on LIBERO, and raises real-world π0.5 success from 38.3% to 73.3%.It also achieves 88.0% on LIBERO-Plus and 61.4% on LIBERO-Pro, and exceeds π0.5 by 3.6 percentage points on LIBERO-Para.Source: §1 · p. 2
• WAVE-0 reports gains under visual, spatial, and language distribution shifts, as well as across three real-world manipulation tasks.The paper also reports greater training efficiency than LaWAM across RoboTwin 2.0 and real-world tasks.Source: §1 · pp. 2–3
• WAVE-0 uses a 257M latent world model to turn VLM-generated action plans into predicted future representations that condition the action expert during training and inference.The model predicts changes in a pretrained VLA's visual representation rather than reconstructing future images.Source: §1 · p. 2, Figure 1 · p. 3

2.1 Vision-Language-Action Policies

• VLA policies combine pretrained vision-language knowledge with robot action supervision, but leave action-dependent environmental dynamics implicit.This motivates integrating explicit future prediction so anticipated action consequences guide action generation.Source: §2.1 · p. 3

2.2 Visual Generation-Based World Action Models

• Video-generation WAMs model future observations and actions but spend computation on large backbones, visual details, and iterative denoising.Some methods retain visual prediction only as training supervision or reduce inference cost through asynchronous execution, sparse interactions, or caching; other visual-generation backbones remain demanding.Source: §2.2 · pp. 3–4

2.3 Latent World Action Models

• Latent world-action methods predict representations without pixel reconstruction, but VLA-JEPA and JEPA-WAM do not directly condition action generation on predicted futures; WAVE-0 does.LaWAM uses visual inverse dynamics to infer latent actions, whereas WAVE-0 predicts transitions from labeled robot action sequences.Source: §2.3 · p. 4

3.1 Problem Formulation

• WAVE-0 plans latent actions, predicts their future world state, and generates actions conditioned on that prediction and the current observation and instruction.The action expert uses both the original context and the predicted future representation.Source: §3.1 · p. 4

3.2 Model Architecture

• Training proceeds in two stages: learn action-conditioned dynamics from ground-truth actions, then integrate the frozen world model into pretrained π0.5 and align VLM-generated plans.Predicted future visual representations condition the action expert alongside the original vision-language context.Source: §3.2 · p. 4, Figure 2 · p. 5
• The latent world model predicts future visual representations from current states and encoded robot actions, using future-representation and residual losses plus counterfactual ranking.Ranking compares demonstrated and mismatched actions against the same observed future.Source: §3.2 · p. 5
• The VLM generates latent action plans aligned with the frozen action encoder’s representations, and those plans condition the frozen world model’s future predictions.The predicted visual consequences are supervised using the future-representation and residual losses.Source: §3.2 · pp. 5–6
• A resampler compresses predicted future representations into tokens that condition the action expert alongside VLM context, while the world model remains frozen in Stage 2.The expert receives planning guidance through predicted visual consequences, not latent-action query tokens.Source: §3.2 · p. 6

4.1 Implementation Details

• The world model uses pooled visual features and a 12-layer Transformer; Stage 1 has approximately 257M trainable parameters, while the full WAVE-0 has approximately 4.3B parameters.The predicted future representations are pooled into conditioning tokens for the action expert.Source: §4.1 · p. 6

4.2 Evaluation on LIBERO

• 98.7% average success on LIBERO, 1.8 percentage points above π0.5; WAVE-0 reaches 97.6% on the Long suite versus 92.4% for π0.5.The benchmark covers four suites and 40 tasks.Source: Table 1 · p. 6, §4.2 · p. 7

4.3 Evaluation on RoboTwin 2.0

• 91.47% average success on RoboTwin 2.0, 11.09 percentage points above π0.5; WAVE-0 reaches 91.46% in clean scenes and 91.48% in randomized scenes.The gains are consistent across both scene settings.Source: Table 1 · p. 6, §4.3 · p. 7

4.4 Out-of-Distribution Generalization

• WAVE-0 achieves 88.0% on LIBERO-Plus, 61.4% on LIBERO-Pro, and 75.0% on LIBERO-Para.The evaluations use a LIBERO-trained model without additional fine-tuning; the supplied passages report the highest compared-method results on Plus and Pro.Source: §4.4 · p. 7, Table 2 · p. 7

4.5 Evaluation on Real-World Tasks

• 73.3% average success across three real-world tasks, versus 38.3% for π0.5 and 43.3% for LaWAM.WAVE-0 also achieves the best performance on each task.Source: §4.5 · p. 8

4.6 Analysis

• 670 H100 GPU hours and 76.8M examples train WAVE-0 on RoboTwin 2.0, versus 1,280 hours and 102.4M examples for LaWAM’s benchmark-specific training.Across three real-world tasks, WAVE-0 processes 33.28M examples versus 345.60M for LaWAM, while achieving higher success rates.Source: §4.6 · p. 8
• WAVE-0’s illustrated attention is more localized around manipulated objects and targets than LaWAM’s, which is more dispersed across robot arms and background regions.Figure 3(b) compares attention from action queries to current main-view image patches at the final denoising step.Source: §4.6 · p. 8

4.7.1 Effects of World Model Integration Strategy

• 91.47% average success with WAVE-0 exceeds the 89.51% achieved by auxiliary supervision alone, with further gains over π0.5 + WM in clean and randomized settings.WAVE-0 explicitly conditions action generation on predicted futures during training and inference; both variants use the same frozen Stage 1 world model.Source: §4.7.1 · p. 9

4.7.2 Effects of Training Objectives

• 0.1648 action sensitivity gap and 89.5% margin satisfaction with ranking supervision exceed 0.0020 and 0% without it at 20k steps.Ranking supervision also improves success by 2.6 and 3.0 percentage points on LIBERO-Plus and LIBERO-Pro, respectively.Source: §4.7.2 · p. 9, Figure 4 · p. 9
• 88.0% on LIBERO-Plus and 61.4% on LIBERO-Pro follow Stage 2 additions of Lworld and Lalign, which provide complementary gains.Adding Lworld improves success by 0.7 and 1.7 percentage points; adding Lalign yields further gains of 2.7 and 1.7 points.Source: §4.7.2 · p. 9, Figure 4 · p. 9

5 Conclusion

• WAVE-0 translates VLM-generated latent action plans into predicted future visual representations that condition the action expert during training and inference.It avoids generating future frames or using predicted futures only as auxiliary targets.Source: §5 · p. 9
• Ground-truth robot actions and counterfactual ranking supervision train the world model to represent action-dependent dynamics and improve policy generalization.Source: §5 · p. 9
• WAVE-0 improves performance on simulation and real-world benchmarks without large-scale world-model pretraining, suggesting an efficient path toward world-model-augmented robot policies.Source: §5 · p. 9

A.1 Training and Evaluation Details on Simulated Benchmarks

• LIBERO training uses a 20k-step latent-world-model stage followed by 10k policy-training steps with the world model frozen; the stages take approximately 10 and 13 hours.Training uses eight GPUs, two camera views, a 7-dimensional action space, and an 8-dimensional proprioceptive state.Source: §A.1 · p. 12
• RoboTwin 2.0 training uses 50k Stage 1 steps and 100k Stage 2 steps with the world model frozen, taking approximately 6 and 78 hours.Observations include three camera views and 14-dimensional proprioception, with a 14-dimensional action space.Source: §A.1 · p. 13
• RoboTwin evaluation averages success over 100 episodes per task in clean and randomized settings, then takes the mean across both settings.All 50 tasks are evaluated.Source: §A.1 · p. 13

A.2 Experimental Details on Real-World Tasks

• The real-robot evaluation covers Book Organization, USB Storage, and Keyboard Pushback, with 200 demonstrations per task and representative WAVE-0 rollouts in Figure 5.Each task-specific policy is trained on its corresponding demonstrations; evaluation uses a dual-arm Piper robot.Source: §A.2 · p. 13, Figure 5 · p. 14
• For each task, π0.5 is fine-tuned for 30k steps on 200 demonstrations and predicts 32-step action chunks, executing 10 actions before replanning.Source: §A.2 · p. 14
• WAVE-0 pretrains its world model for 40k steps on all 600 demonstrations, then freezes it while training each task-specific policy for 30k steps.The world model predicts representations 32 steps ahead, and the policy generates 32-step action chunks using 32 latent action queries.Source: §A.2 · p. 14
• LaWAM trains for 100k steps and returns 18 actions from a 50-step horizon, whereas ImageWAM trains for 60k steps and outputs 16-step action chunks.Source: §A.2 · pp. 14–15
• 305.9 ms mean latency for WAVE-0 is 18.8 ms above π0.5 and 47.5% lower than ImageWAM; LaWAM’s latency is 223.9 ms.LaWAM is smaller (2.3B versus 4.3B parameters) and returns 18 rather than 32 actions per request, factors to consider in the comparison.Source: §A.2 · p. 15, Table 11 · p. 16

A.3 LIBERO-Pro Detailed Results

• Table 12 reports LIBERO-Pro results across different perturbation categories.Source: Table 12 · p. 16

A.4 RoboTwin Detailed Results

• Table 13 reports per-task success rates for RoboTwin 2.0.Source: Table 13 · p. 17