Real-time robot control has a built-in dilemma: to infer object motion and task progress, a policy needs a long visual history — but the longer the history, the slower inference, and actions fall behind the physical world. Long-WAM, submitted to arXiv on October 7 by an NVIDIA research team (arXiv:2610.10528), attacks both sides at once: it scales a world-action model's visual context to 19.2 seconds while a co-designed runtime keeps each action chunk in the 100-ms class.

Longer memory helps — if pretraining is autoregressive

The paper's central finding: access to history is not the same as using it. The video expert continues autoregressive training from the LongLive2.0-Robot checkpoint on roughly 10,000 window-equivalent hours of robot and egocentric video (RoVid-X, AgiBot World, EgoDex, EgoVerse, VITRA), with no action labels, predicting the next chunk on sequences up to 30 seconds. It is then coupled to an action expert through an asymmetric attention interface (a Mixture of Transformers); video queries never read action tokens.

The payoff is in the numbers. On RoboCasa GR-1, scaling observed context from 2.4 s to 19.2 s lifts success from 66.3% to 78.7%; on LIBERO-Long, 2.4 s of history already lifts success from 94.5% to 99.5%. The control group is the telling part: with a bidirectionally pretrained Wan2.2 initialization, the same 19.2 s yields only 61.6% — the project page calls it "no net gain." In other words, the long-context dividend belongs to autoregressive pretraining. The boundary is stated honestly: performance dips at 38.4 s, but 80.4% of sampled history frames there are padding (training trajectories average 12.1 s), so the dip reflects data, not a memory limit.

Imagine the future in latent space, then act

Inference runs in an inverse-dynamics (IDM) mode: the video expert predicts partially denoised future latents from history without decoding pixels; the action expert denoises an action chunk conditioned on history plus the predicted future. Real-robot results stand out: on a Unitree G1, grasping cups off a moving conveyor succeeds 100/100/95/90% at 3.0/4.5/6.0/7.5 cm/s, while both baselines fail every trial at 6.0 and 7.5 cm/s. Dynamic cup stacking lands 19/20 (95%) versus 0/20 for both Pi0.5 and Fast-WAM, with human teleoperation at 50/54 (92.6%). On the YAM platform, tasks longer than 40 s average 81.7%. Paired with GPT-6 Astra as a high-level planner — same checkpoint, no further training — overall success on RoboCasa365 rises from 31.4% to 54.4%, and unseen compositions jump from 6.1% to 35.0%.

107.4 ms: real-time is a systems problem

A clever model that misses its deadline is useless. Streaming VAE encoding, pure asynchronous execution, NVFP4 quantization, denoising-invariant KV reuse, CUDA Graphs and device-specific kernel tuning compress the full predict-then-act pipeline to 107.4 ms per action chunk on an RTX 5090 (future-video latent prediction included), a 3.3x speedup over BF16 eager; DGX Spark runs at 328.2 ms (4.1x) and Jetson AGX Thor at 378.7 ms (3.2x). For latency context: LingBot-VA takes 3618.4 ms, Motus 1201.1 ms, Fast-WAM 244.1 ms. Under pure asynchronous execution, RoboTwin 2.0 success is 94.2% versus 94.4% synchronous — essentially lossless.

So what

Long-WAM matters beyond robotics: it is a controlled experiment showing that context gains depend on the pretraining paradigm — the long-context story the LLM world has been racing through for two years now has its quantified embodied counterpart. For anyone building VLA policies, choosing an autoregressive versus bidirectional video foundation may matter more than another order of magnitude of data. Paper and full numbers: project page (https://nvlabs.github.io/LongLive/Long-WAM/) and arXiv:2610.10528 (https://arxiv.org/abs/2610.10528).