OpenWAM: An Open Framework for Composable World-Action Models

Researchers have introduced OpenWAM, an open framework for composable world-action models (WAMs) designed to decouple video backbones, interaction structures, and supervision methods. By providing a shared causal robot-video foundation and configurable video-action interaction programs, OpenWAM addresses the difficulty of comparing design choices in existing WAMs that simultaneously alter multiple architectural components.

Architecture and Pretraining

The framework builds upon the Wan2.2-5B model, performing causal robot-video pretraining on approximately 3.34 million trajectories totaling 14.64k hours of source video. This dataset includes real and synthetic robot manipulation, human-guided manipulation, and human interaction from sources like Open X-Embodiment, AgiBot World Beta, and InternData-A1. Training was conducted on 32 NVIDIA B200 GPUs for 14 days.

Key architectural features include:

• Mixture-of-Transformers (MoT): Combines a 5B video expert and a 2B action expert, initialized from width-adapted copies of pretrained video layers.
• Causal Attention: Generates video in chunks, where each chunk references only current/past observations and previous chunks.
• Four Interaction Programs: Supports Video-then-action (VTA), Action-then-video (ATV), Joint, and Decoupled generation orders using latent flow matching.

Performance and Benchmarks

OpenWAM demonstrates strong performance in both simulation and real-world settings:

• LIBERO Benchmark: The OpenWAM-VTA variant achieved a mean success rate of 98.6% (Object: 99.4%, Goal: 98.4%, Spatial: 98.6%, Long: 97.8%). This outperforms baselines such as LingBot-VA (98.5%) and Motus (97.7%).
• Real-World Manipulation: On bimanual tasks using two Franka Research 3 arms, OpenWAM-VTA achieved a mean success rate of 92.1% across toasting bread, solving a Rubik's cube final layer, and sorting cups.
• Ablation Results: Robot-video pretraining improved LIBERO-Long success by 29.4 points for VTA and 34.4 points for Joint compared to original Wan2.2 initialization. The MoT architecture provided a 5.0-point improvement over shared DiT for VTA.

Counterfactual Data and Dynamics Models

The study introduces LIBERO-Long-CF, a dataset of 32,000 counterfactual segments generated by simulating alternative action sequences, including failures. This data significantly enhances the reusability of independently trained dynamics components:

• Inverse Dynamics Model (IDM): Using counterfactual data, the local-context IDM achieved 84.0% mean success on held-out LIBERO-90 tasks, compared to 21.5% with demonstration-only data.
• Forward Dynamics Model (FDM): Counterfactual supervision reduced RGB prediction error by 34.5% and improved the ability to identify correct future outcomes from 16 alternatives from 21.1% to 71.3%.

Unified vs. Specialized Models

A unified OpenWAM checkpoint supporting policy, IDM, and FDM objectives achieved 92.8% success on LIBERO-Long, surpassing the released UVA system (88.0%). However, specialized models still offer higher task success and more accurate counterfactual predictions, suggesting a trade-off between composability and peak performance.