Daily AI Paper

Fast robot control has a deceptively hard requirement: the robot needs enough visual history to understand what is changing, but the more history you process, the more delay you introduce before the next action. And in closed-loop control, delay is not a small inconvenience. It can be the difference between catching motion, missing a grasp, or acting on stale information. This paper is about that tension, and about a bigger question behind it: when does longer context actually help a world-action model, and when is it just extra computation? To understand the setting, it helps to know what world-action models are trying to do. These are generative robot policies that connect video understanding with action generation. Instead of only mapping the current image to the next move, they try to model how the scene will evolve, then use that prediction to choose actions. That matters because a single frame often hides the thing that really matters for control: motion, interaction progress, and whether an object is about to move, slip, or be grasped. Visual history can resolve that ambiguity. But most prior systems have treated history as something to store or retrieve, not necessarily something the model has truly learned to exploit. That distinction is central here. The authors argue that access to history is not the same as using it. If a model was pretrained in a way that already learned causal, history-to-future prediction, then longer context may become genuinely useful. If it was pretrained bidirectionally, the same extra history may not translate into better control. So Long-WAM asks not just how much context a robot policy can ingest, but how the pretraining recipe determines whether that context pays off. Their approach has two main parts: a predictive foundation, and a deployment system that can use it in real time. On the modeling side, they first train a video foundation model in an autoregressive way on robot and egocentric video, without action labels. The idea is to learn physical dynamics from observation alone: how objects move, how interactions unfold, and how future frames follow from past frames. This foundation, called LongLive2.0-Robot, continues training from a prior AR checkpoint using about 10,000 window-equivalent hours of robot and egocentric video from several datasets. Because this is video-only pretraining, it can learn across different embodiments without needing a shared action space. A key detail is that the pretraining is causal. The model sees history and learns to predict the future in order. That matters because the downstream robot policy is also causal. During adaptation, the video expert and the action expert are coupled asymmetrically: video queries can only look at their own past visual blocks, while action queries can attend to observed history, partially denoised future latents, and the noisy action chunk. In other words, the model first imagines what is likely to happen next, and then uses that imagined future to decide the action. The paper describes this as an inverse-dynamics-style predict-then-act setup. The adaptation is trained in two passes. First, there is video flow matching conditioned on clean history. Then there is action flow matching conditioned on history plus a forward-noised ground-truth future. Importantly, the future latents are not decoded back into pixels at inference time. They remain latent, and their cache is reused during action denoising. That keeps the predictive conditioning available while avoiding extra overhead. The authors also preserve the causal temporal structure from the pretrained video model instead of replacing it during robot adaptation. The paper then studies context scaling directly. They vary the duration of real observed history in the causal prefix, while keeping the prediction horizon and action horizon fixed within each benchmark. So the question is not whether the model can plan farther ahead, but whether more past evidence improves the same near-term control decision. They compare three initializations: a bidirectionally pretrained model, a general autoregressive video model, and a robot-domain autoregressive model. This lets them isolate whether longer history helps only when the foundation already learned to predict causally. There is also a systems contribution, because longer context is only useful if the robot still responds on time. To make that possible, Long-WAM combines asynchronous execution, streaming observation encoding, low-precision acceleration, and device-specific kernel tuning. The system overlaps inference with robot execution, so the next action chunk is prepared while the current one is being executed. It also encodes incoming observations in a streaming fashion, reducing post-trigger work. With NVFP4 quantization and other optimizations, the pipeline runs on hardware such as RTX 5090, DGX Spark, and Jetson AGX Thor. On RTX 5090, each action chunk, including the future-video latent prediction, takes 107.4 milliseconds, which the paper reports as a 3.3x speedup over BF16 eager execution. The experimental picture is fairly broad. On RoboCasa GR-1 Tabletop, increasing the history window from 0.0 to 19.2 seconds raises success from 63.3 percent to 78.7 percent. That is a large gain, and it is the paper’s clearest evidence that longer context can matter a lot. But the effect depends on the initialization. The autoregressive foundations benefit from more history, while the bidirectionally pretrained initialization shows no net gain. The gap between the robot-domain AR model and the bidirectional one also widens as history grows, suggesting that the causal pretraining is what makes long context useful rather than merely available. On LIBERO-Long, 2.4 seconds of history improves success from 94.5 percent to 99.5 percent. The gains are smaller in absolute terms because the baseline is already high, but they still support the same story: when the model has learned causal prediction beforehand, extra context can improve control. The robot-domain AR pretraining further raises peak success on both LIBERO-Long and GR-1. The paper also reports strong benchmark results beyond these context studies. Long-WAM achieves the best results among compared methods on LIBERO-Long, RoboTwin 2.0, and DOMINO. The reported average success is 94.4 percent on RoboTwin 2.0 and 34.9 percent on DOMINO. Those numbers are useful because they show the system is not only good in one carefully chosen setup; it is competitive across several manipulation benchmarks. The real-time deployment results are especially striking. On a Unitree G1 robot, Long-WAM grasps cups from a conveyor moving at 7.5 centimeters per second in 90 percent of trials, and stacks moving cups in 95 percent of trials. The paper emphasizes that neither Pi0.5 nor Fast-WAM succeeds in any of 20 trials in those dynamic cup-stacking settings. On YAM, Long-WAM completes tasks lasting more than 40 seconds on average with 81.7 percent success, which extends the evaluation from quick interception to sustained long-horizon execution. So this is not just about reacting faster; it is about maintaining coherent behavior over time. There is also a hierarchical planning result that helps clarify where Long-WAM fits. On RoboCasa365, pairing the unchanged Long-WAM checkpoint with a higher-level planner, GPT-6 Astra, raises overall success from 31.4 percent to 54.4 percent, compared with 25.2 percent for the planner alone. The gain from adding the planner is larger for Long-WAM than for Pi0.5. That suggests a useful division of labor: a stronger executor makes planning more valuable, because the planner can rely on the low-level policy to carry out more complex instructions faithfully. So what should we take away? The main message is that context is a resource, but its value depends on how the model was pretrained and whether the system can use that context without becoming too slow. If the foundation learned causal history-to-future prediction during pretraining, then longer observation windows can significantly improve robot control. If it did not, simply adding more history may not help. Long-WAM’s second message is that these modeling gains only matter if the deployment stack is engineered carefully enough to preserve real-time responsiveness. There are also some limitations and open questions. The paper’s strongest gains come from specific combinations of autoregressive pretraining, causal adaptation, and optimized deployment, so it remains to be seen how broadly the same recipe transfers to other robot embodiments and task families. The context studies are benchmarked at fixed horizons, which is good for isolating the effect of history, but it does not answer every question about longer-term planning or memory beyond the tested windows. And while the asynchronous system is impressive, it depends on specialized acceleration and hardware-aware tuning, so portability to less capable devices may be more challenging. Still, the broader lesson is clear. For generative robot control, history is not just something to store. It is something the model must be trained to understand causally. Long-WAM shows that when you get that pretraining right, and when you engineer the system so the robot can act in time, longer context can become a real advantage rather than a latency tax.