OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining

• OpenWAM-Infra factorizes WAM design choices into modular components and unifies training, inference, deployment, and evaluation across domains and embodiments.The infrastructure enables controlled comparison across WAM designs.Source: §1 · pp. 3–4
• Upstream world knowledge transfers most effectively through a sufficiently capable generative backbone and a compact, information-rich visual latent space.Source: §1 · p. 3
• World–action synergy requires sufficient action-specific capacity, explicit world-to-action information flow, and a joint test-time denoising schedule—not parameter count or joint prediction alone.Source: §1 · p. 3
• Embodied pretraining principally improves out-of-domain generalization; human egocentric video broadens world coverage, robot trajectories ground executable actions, and joint training integrates both.The study identifies joint training as the strongest practical integration strategy in its experiments.Source: §1 · p. 3
• OpenWAM-α, pretrained on 518M frames (≈6,400 hours) through an 80-D action space, performs strongly across eight simulation benchmarks and real-robot single-arm, bimanual, and dexterous-hand platforms.The evaluations span five embodiment categories in simulation and three platform types in real-robot experiments.Source: §1 · p. 4
• OpenWAM releases its infrastructure, evaluation protocols, pretrained weights, and data recipes to support reproduction and further research.Source: §1 · p. 4

World–Action Models

• WAMs connect predictive world models to executable behavior and inherit video-generative visual-dynamics priors, unlike VLAs’ vision–language knowledge; existing WAM designs remain coupled.World models also support planning, model-based reinforcement learning, and policy evaluation.Source: §2 · p. 4

Open Research Ecosystems for Generalist Robot Policy Learning

• OpenVLA, StarVLA, StarVLA-α, and XPolicyLab advance open VLA models, modularity, and evaluation, while comparable open research ecosystems remain largely absent for WAMs.The passage argues that a modular WAM system paired with a strong pretrained model could democratize research and support principled study and scaling.Source: §2 · p. 4

Towards a Scientific Understanding of Model Design

• OpenWAM brings controlled empirical design to WAMs through OpenWAM-Infra, OpenWAM-Study, and OpenWAM-α, addressing inheritance, synergy, and scaling under heterogeneous multi-domain pretraining.This builds on research that decomposes model design into controlled variables, tests mechanisms, derives recipes, and validates whether they persist at scale.Source: §2 · pp. 4–5

3.1 Composable Model

• OpenWAM-Infra composes models from a visual encoder, stream backbones, and visibility masks that govern information flow within and across streams.The encoder maps observations to latents, while backbones process world, action, and optional understanding streams.Source: §3.1 · pp. 5–6
• Its parameter-free composition rule supports shared-parameter single-system, separate-backbone dual-system, and VLM-augmented tri-system architectures.Composition sequences backbone operations and can split layer steps to insert joint attention without modifying backbone internals.Source: §3.1 · p. 7

3.2 Training Runtime

• One architecture-agnostic trainer applies a joint flow-matching objective to video and action predictions, independently noising each stream across the full noise-timestep plane.Self-contained checkpoints preserve model weights, resolved module configuration, and action-normalization statistics for fine-tuning, resumption, and deployment.Source: §3.2 · pp. 7–8

3.3 Deployment Runtime

• A single policy server separates synchronous or asynchronous inference modes from denoising schedules, including synchronized and stream-leading paths.Independent timestep sampling during training keeps asynchronous schedules in-distribution, and checkpointed timestep warps keep training and serving noise grids aligned.Source: §3.3 · pp. 8–9

3.4 Evaluation Protocol

• Benchmarks connect to the policy server as thin clients, supporting eight simulation benchmarks and real-robot platforms through the same interface.A unified 80-D action space maps embodiment-specific dimensions into fixed slots, while server-side conversions return actions in native physical units.Source: §3.4 · pp. 9–10
• The validity mask marks mapped slots, preventing gradients on unmapped coordinates during training while leaving them on their analytic noise path at inference.Incoming proprioception traverses the same map in the forward direction.Source: §3.4 · p. 10

4 OpenWAM-Study: Design Principles for World–Action Models

• The study evaluates success rate on RoboTwin2.0-Full and Clean2Random, where clean-only training followed by randomized-scene evaluation tests out-of-domain generalization.RoboTwin2.0 is described as a bimanual manipulation benchmark spanning over 50 tasks.Source: §4 · p. 11

4.1.1 Generative World Priors

• WAM success improves with larger video-generation backbones; Wan2.1-I2V-14B performs best, but Wan2.2-TI2V-5B trails by only 1.40 points and is adopted for efficiency.The comparison also varies backbone architecture, pretraining data, and objective, so it does not attribute the entire gain to parameter count alone.Source: §4.1.1 · p. 11, Figure 6 · p. 11

4.1.2 Visual Representation Priors

• Compact, information-rich latents matter more than encoder category: representation encoders with S-VAE dimension contraction outperform FLUX.2-VAE, and DINOv3 w/ SVAE approaches Wan2.2-VAE.Naively adopting high-dimensional representation latents performs worse; temporal compression and dimension contraction enable their effective use.Source: §4.1.2 · pp. 12–13, Figure 7 · p. 12

4.2.1 Architectural Capacity

• Performance rises with architectural capacity: tri-system achieves the best overall result, while dual-system joint self-attention is selected to balance performance and complexity.The dual-system choice also isolates world-knowledge/action-learning interaction from potential influence of VLM understanding features.Source: §4.2.1 · p. 13, Table 1 · p. 13

4.2.2 Training-Time Information Flow; 4.2.3 Inference-Time Information Flow

• World–action synergy requires explicit world-to-action information flow during training and synchronized joint denoising at inference; no tested asynchronous schedule improves performance.Synchronized denoising performs best regardless of which stream leads or how relative progress is parameterized.Source: §4.2.2 · pp. 13–14, §4.2.3 · p. 14, Figure 9 · p. 14

4.3.2 Embodied Pretraining Primarily Expands OOD Generalization

• Embodied pretraining mainly improves out-of-domain generalization; robot-only data favors in-domain performance, while mixed sources generalize better and one-stage co-training is a practical default.Sequential training and one-stage co-training perform nearly equally in-domain and out-of-domain, with co-training marginally strongest overall.Source: §4.3.2 · p. 15

4.3.3 Pretraining Changes the Preferred Information Flow

• After embodied pretraining, mutual world–action visibility is preferred across evaluation protocols, enabling predicted frames to guide actions and predicted actions to inform frame synthesis.Without pretraining, RoboTwin2.0-Full slightly favors one-way visibility; the preference reverses after pretraining, including on both Clean2Random splits.Source: §4.3.3 · p. 16, Figure 11 · p. 16

5.1.1 Architecture

• The architecture pairs a frozen Wan2.2-VAE and pretrained Wan2.2-TI2V-5B world stream with a dedicated 1B-parameter ActionDiT, jointly denoising future video and 80-D actions through mutual-attention bridge layers.The streams are conditioned on the current observation, language, and proprioception, and use separate noise timesteps.Source: Figure 12 · p. 17, §5.1.1 · p. 18

5.1.2 Training; 5.1.3 Deployment

• Training uses one-stage end-to-end co-training on egocentric and robot data, while inference synchronously denoises both streams along the joint-noise diagonal in 10 steps.Training draws stream timesteps independently; evaluation pauses whenever the action buffer empties and executes each predicted chunk as produced. The 10-step loop takes roughly 170 ms per chunk on an RTX 5090.Source: §5.1.2 · p. 18, §5.1.3 · pp. 18–19

5.2.1 Pretraining Data Mixture

• Pretraining uses 518M frames (about 6,400 hours) from egocentric human, real-world robot, and synthetic robot data, combining broad visual coverage with grounded robot actions and varied manipulation skills.Egocentric data has no robot action labels and supervises only the world stream; real-robot data supplies executable trajectories, while synthetic data broadens single-arm and bimanual skill coverage.Source: §5.2 · p. 19, §5.2.1 · p. 19

5.2.2 Data Curation

• Data curation applies shared visual-quality screening and robot-specific signal cleaning, removing defective video, invalid egocentric segments, misaligned or unreliable signals, idle footage, and severely degraded episodes.Robot episodes are discarded for video–signal misalignment affecting more than 2% of frames or when over 70% of frames are removed; static leading and trailing segments are trimmed, but mid-episode pauses are retained.Source: §5.2.2 · p. 20

5.3.1 How Does OpenWAM-α Perform?

• Across simulation benchmarks, OpenWAM-α is top-tier on five, leads EBench by roughly 4 points in both SR and Score, and is the strongest WAM on RoboTwin2.0-Clean2Random and RoboDojo.It remains within a marginal gap of the best model on LIBERO, VLABench, RoboTwin2.0-Full, RoboCasa365, and RoboCasa-GR1.Source: Figure 14 · p. 21, §5.3.1 · p. 21
• On LIBERO-Plus, OpenWAM-α falls below its standing elsewhere, particularly under camera and noise perturbations, as limited single-arm data compounds pixel-latent sensitivity to visual shifts.The paper attributes its limited single-arm coverage to DROID’s fixed platform and viewpoints plus the single-arm portion of InternData-A1; it contrasts pixel reconstruction with more robust semantic-level features.Source: §5.3.1 · pp. 22–23

5.3.2 VLA versus WAM: Which Paradigm Prevails?

• VLA and WAM groups show no substantial overall success-rate gap: WAMs fit in-distribution tasks better through video-latent supervision, while VLAs generalize better out of distribution as WAM rollout errors accumulate.The reported takeaway is that neither paradigm prevails outright and sufficiently large, diverse pretraining data can compensate for either deficit.Source: §5.3.2 · pp. 23–24, Figure 16 · p. 24

5.4 Real-Robot Evaluation

• Real-robot evaluation examines OpenWAM-α’s general capability and generalization across single-arm, bimanual, and dexterous-hand embodiments.The supplied passage introduces the evaluation setups but does not report their results.Source: §5.4 · p. 24
• On single-arm tasks, OpenWAM-α leads LingBot-VA and π0.5 on most tasks and achieves the best average success rate.The experiments use the Franka-Research-3 platform and assess stacking, pick-and-place, and hanging.Source: §5.4 · pp. 24–25, Table 6 · p. 25
• OpenWAM-α tops the RoboDojo bimanual leaderboard and reaches state of the art on most tasks across three embodiments.It remains consistently strong across the embodiments and their tasks.Source: §5.4 · p. 25, Table 7 · p. 26
• On the unseen Wuji-hand and Tianji-arm embodiment, OpenWAM-α outperforms π0.5 across all tasks and in-domain and out-of-domain setups.The platform and its action space—a 9-D end-effector pose combined with 21 dexterous-hand degrees of freedom—were absent from pretraining.Source: §5.4 · pp. 25–26, Table 8 · p. 26
• The real-robot evaluation spans single-arm, bimanual, and dexterous-hand manipulation, and OpenWAM-α performs strongly in every setting.The bimanual experiments span ARX X5, Piper, and Piper X; dexterous-hand tests use the Wuji hand and Tianji arm. Evaluations report success rate and, for bimanual and dexterous tasks, Progress Score.Source: §5.4 · pp. 24–26

6 Conclusions

• OpenWAM-α applies the distilled recipe to egocentric human and robot data through a unified action space, with strong results across simulation and real-robot experiments.Evaluations span single-arm, bimanual, and dexterous-hand platforms.Source: §6 · p. 26
• OpenWAM-Infra factorizes the WAM design space into composable modules and unifies training, policy serving, and evaluation across eight simulation benchmarks and real robots.The modules form three architecture families.Source: §6 · p. 26
• OpenWAM-Study examines inherited world knowledge, world–action learning synergy, and its consolidation across domains, distilling the findings into a concrete recipe.Source: §6 · p. 26
• The authors release infrastructure, evaluation protocols, pretrained weights, and data recipes as a reproducible foundation, with OpenWAM-α as a baseline for future comparison.Source: §6 · p. 27
• Future directions include larger, more diverse embodied datasets with precise action annotations, compact and robust visual representations, and end-to-end designs combining WAM and VLA strengths.Section A provides a more detailed discussion of limitations and future work.Source: §6 · p. 27, Appendix · p. 33

Appendix

• The appendix supplements the main paper with analyses and implementation details, including pretraining and post-training configurations, real-world task setups and protocols, and full per-benchmark simulation scores behind Figure 13.Source: Appendix · p. 33

4. Visual encoder

• A compact, information-rich latent remains needed: Wan2.2-VAE is a current trade-off, as pixel-reconstruction encoders lack robustness to viewpoint, noise, and scene variation.The paper calls for deeper exploration of a latent representation that preserves sufficient environment information while remaining compact.Source: A Limitations and Future Work · p. 33

3. Pretraining data mixture

• OpenWAM omits UMI-style data (UMI (Chi et al., 2024), DexUMI (Xu et al., 2025)); diverse executable robot-action data and simulation could expand scene and task coverage.The authors suggest UMI-style interfaces or post-processing pipelines and identify simulation as a way to scale robot data beyond physical-world time and labor constraints.Source: A Limitations and Future Work · p. 33

1. Training phases

• OpenWAM focuses on embodied pretraining; post-training and adaptation could substantially improve embodied foundation models, but their recipes and mechanisms remain open.Source: A Limitations and Future Work · p. 33

2. Architecture

• Across six supported architecture variants, OpenWAM mainly studies cross-modality attention and hard-routed MoE, leaving early-fusion tokenization and soft-routed MoE for future exploration.Source: A Limitations and Future Work · p. 33

B Training Details

• Training details are organized into multi-domain pretraining and dataset-specific supervised fine-tuning during post-training.Source: B Training Details · p. 33

B.1 Pretraining Configuration

• Pretraining used 128 H200 GPUs for approximately seven days, with a global batch size of 3,072 clips per optimizer step.The setup used 16 nodes with eight GPUs each, 24 clips per GPU, and no gradient accumulation; source-level data budgets and realized mixture proportions are reported separately in Table 4.Source: §B.1 · p. 33, Table 9 · p. 34

B.2 Dataset-Specific SFT Configuration

• All downstream models start from the same pretrained OpenWAM-α checkpoint, and SFT retains the pretraining configuration except for three benchmark-specific settings.Those settings are global batch size, training length, and whether image augmentation is applied.Source: §B.2 · p. 34
• When enabled, image augmentation uses the same ColorJitter(0.2, 0.2, 0.2, 0.0) as pretraining.Source: §B.2 · p. 34
• LIBERO-Plus is evaluated with the LIBERO checkpoint and receives no further fine-tuning.Source: §B.2 · p. 34

• For each embodiment, the section specifies the task setup and corresponding protocol for the real-world evaluation in Section 5.4.Source: C Real-World Evaluation Protocols · p. 35

C.1 Single-Arm Real-Robot Experiments

• Six real-world tabletop tasks cover stacking, hanging, and drawer manipulation with a Franka-Research-3 arm and parallel gripper.Each task uses a natural-language instruction and a fixed physical scene.Source: §C.1 · p. 35, Figure 18 · p. 35
• The Franka-Research-3 setup uses RGB observation cameras for closed-loop execution across the six tasks.The workspace includes the drawer, stacking objects, and hanging fixtures; the robot is equipped with an Intel RealSense camera and a Robotiq parallel gripper.Source: Figure 18 · p. 35, §C.1 · p. 35
• Figure 19 visualizes each task’s execution as six uniformly sampled frames from a rollout video.Source: §C.1 · p. 35, Figure 19 · p. 36
• Each task uses 100 task-specific real-robot demonstrations for fine-tuning and is evaluated over 20 independent trials.Success is determined by task-specific completion criteria: stable stacking, remaining hung on the shelf, or placing the object inside and closing the drawer.Source: §C.1 · p. 36

C.2 Dexterous-Hand Real-Robot Experiments

• The physical dexterous-hand robot autonomously executes natural-language instructions on Stack Toy Tower, Collect Shuttlecocks, Twist off Bottle Cap, and Put Away Clothes.The platform operates in the corresponding task scene; the setup includes a workspace and observation camera.Source: §C.2 · p. 37
• Each task is tested across multiple independent trials under ID and OOD conditions, with Score measuring completed manipulation elements and SR measuring complete-task success.OOD conditions vary object identity, layout, illumination, or background appearance one factor at a time while preserving the task instruction and overall objective.Source: §C.2 · pp. 37–38
• Success requires completing each full task objective, while progress reflects partial completion: placing shuttlecocks, stacking all three discs in descending size order, putting clothes away, or twisting off the cap with a stable grasp.For bottle-cap trials, progress records whether the cap is twisted off even if the final stable grasp is not achieved.Source: §C.2 · pp. 38–39

C.3 Bimanual Real-Robot Experiments

• The evaluation covers all 18 tasks of RoboDojo-Real on its official real-robot evaluation platform.Source: §C.3 · p. 39
• The platform evaluation spans three bimanual embodiments: ARX X5, Piper, and Piper X.Source: §C.3 · p. 39
• The evaluation strictly follows the unified protocol defined by the official evaluation team.Source: §C.3 · p. 39

D Per-Benchmark Simulation Results

• The tables report full scores for each simulation benchmark except LIBERO-Plus, whose scores appear in main-text Table 5.Figure 13 summarizes the tabled results.Source: D Per-Benchmark Simulation Results · p. 39
• Within each table, baselines are grouped as VLA or WAM; bold marks the best value and underline the second best.Source: D Per-Benchmark Simulation Results · p. 39