Reproduce “Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models” on OpenResearch

About this paper

Weiheng Zhao, Haoyi Jiang, Xin Shi, Liu Liu, Fan Huang, Zhizhong Su, Wei Sui, Xinggang Wang

Faster-WAM introduces an efficient architecture for Vision-Language-Action models that explicitly integrates future visual information during inference for robust robot manipulation. It achieves a 2.21x speedup over prior robust World Action Models (WAMs) while improving average success rates to 73.6% on LIBERO-Plus out-of-distribution tasks and 71.1% in real-world OOD scenarios.

Key insights

• Inference-time future conditioning is critical for the robustness and generalization of World Action Models, particularly under out-of-distribution settings, rather than merely serving as a training signal.
• Efficient integration of future visual context into action policies can be achieved by generating it once and reusing it sparsely, avoiding the computational cost of dense, iterative video-action denoising.
• Sparse interaction and multi-depth information fusion mechanisms (SparseMoT and Interval KV-Fusion) enable a superior balance between robustness and inference efficiency in WAMs.

Summary by alphaXiv.