OpenWAM, an Open-Source Stack for Comparing Robot WAM Designs, Released

Researchers from the National University of Singapore, Tsinghua University, and other institutions released the OpenWAM stack, which modularizes design choices to isolate performance drivers in robot World-Action Models (WAM). It was developed to address the difficulty of assessing the contribution of individual elements in existing WAMs, where backbone, representation, and inference are coupled.

OpenWAM Structure and Features

OpenWAM consists of OpenWAM-Infra (modular infrastructure), OpenWAM-Study (controlled experiments), and OpenWAM-α (open model). It is based on PyTorch and supports Hydra YAML configuration and WebSocket deployment.

• Modularity: Allows experimentation by combining 4 visual encoders, 5 video backbones, 6 architectures, 4 attention masks, and more.
• Architecture: Supports Single-system, Dual-system, and Tri-system (which adds a VLM understanding expert).
• Inference Acceleration: Improves inference speed by applying DiT cache and torch.compile.

OpenWAM-α Model and Performance

OpenWAM-α was pre-trained on approximately 6,400 hours of data including human videos and robot data. It uses Wan2.2-VAE as the visual encoder and Wan2.2-TI2V-5B DiT as the world stream.

• WAM vs VLA: WAM uses both action supervision and future video latent prediction, resulting in high accuracy on ID (in-distribution) tasks but inferior performance on OOD (out-of-distribution) tasks due to error accumulation over long prediction horizons. In contrast, VLA uses only action supervision, showing relative advantage in OOD scenarios.
• Benchmarks: Showed vulnerability to camera and noise perturbations on benchmarks such as LIBERO-Plus. This is analyzed as a limitation of pixel-level prediction.

Application Guide and Requirements

• Recommended for: Architecture comparison researchers, fine-tuning teams.
• Not recommended for: Teams seeking immediate deployment (requires 8 GPUs with 80GB VRAM in the recommended environment, and is vulnerable to perturbations).
• License: OpenWAM itself follows Apache-2.0, while video backbones and benchmark data are subject to separate licenses.