Rolling-WAM: World Action Models with Rolling Imagination

What it claims

Rolling-WAM cuts closed-loop replanning latency for World Action Models by never re-denoising the full prediction horizon from scratch. The page puts concrete numbers on the tradeoff the summary only gestures at: measured inference latency drops from 978 ms/replan (Joint-WAM) and 548 ms (Fast-WAM) to 215 ms (Rolling-WAM) — the source of the 4.5× claim — while real-world success on Unitree G1 tasks rises (78.3% Joint-WAM, 75.0% Fast-WAM, 85.0% Rolling-WAM) and simulation success is 93.3% (vs 90.6/91.9). LIBERO average is reported at 98.1%, against baselines π0 (94.1) and π0.5 (96.9).

How it works

The prediction window is split into W temporally-aligned video-action chunks. Two noise schedules run the show:

• Rolling mode (steady state): chunk j sits at noise level σ((j−1+τ)/W). As one denoising cycle completes, the nearest chunk becomes clean and every retained chunk lands exactly at the starting noise level for its new slot — the key identity σ_{j+1}(0)=σ_j(1). A fresh Gaussian-noise chunk is appended to refill the window.
• Initialization mode (episode start): the whole window starts from pure noise, nearer chunks denoise earlier, ending at the rolling-mode configuration.

flowchart LR
  I[Init mode: N steps from full noise] --> R
  R[Rolling mode: N/W steps per cycle] --> E[Execute chunk 1 - K actions]
  E --> S[Slide window, append Gaussian chunk]
  S --> O[Acquire new observation + state]
  O --> R

The cost saving: each chunk still traverses all N denoising steps, but spread across cycles, so a steady-state cycle needs only N/W steps to produce the next executable chunk. Architecture is a Mixture-of-Transformers pairing a pretrained video DiT with a lightweight action Transformer; masked joint attention lets each action chunk attend to visual features across the entire window, action-to-action attention is restricted to the same chunk, and video tokens never attend to actions. Trained with flow matching, sampling rolling vs init mode per window with probability β.

Limits

The fetched text cuts off mid-results (LIBERO Table I is truncated; RoboTwin per-suite numbers, latency methodology, and the ablations on window size, noise schedule, and cross-chunk attention are named but not shown). The steady-state speedup assumes fixed N and sufficient GPU parallelism to process the enlarged window — larger W lowers latency only under that assumption, so the gain is not free on constrained hardware. No license, weights, or code beyond the project page (rolling-wam.github.io) is stated here.