Efficient-WAM: 1B-parameter robot model runs 30x faster, imagines future in 100ms
Robot AI now predicts futures 30x faster—cost drops without sacrificing action quality.
World-Action Models (WAMs) combine future visual prediction with action generation for embodied AI, but their high inference latency has hindered real-time robot deployment. Researchers from Peking University and other institutions now present Efficient-WAM, a 1B-parameter model that achieves a 30x speedup—reducing per-chunk latency to around 100ms—by treating future video prediction as a compact guidance signal rather than optimizing for visual fidelity.
Efficient-WAM achieves this efficiency through three key innovations: a compact video expert distilled from Wan-2.2-5B, token-sparse video latents that reduce computation, and asymmetric video-action denoising that allocates fewer sampling steps to video than to actions. On RoboTwin 2.0 and real-world manipulation tasks, the model maintains competitive control performance despite visibly coarser future predictions, proving that photorealistic accuracy is not necessary for effective robot action planning.
- Efficient-WAM is a 1B-parameter World-Action Model that achieves ~100ms per-chunk latency, 30x faster than prior WAMs.
- It uses a compact video expert from Wan-2.2-5B, token-sparse latents, and asymmetric denoising to reduce future imagination costs.
- Despite coarser future video predictions, it maintains strong action performance on RoboTwin 2.0 and real-world manipulation tasks.
Why It Matters
Enables real-time robot deployment with future-aware planning, cutting compute costs dramatically for embodied AI systems.