Why more history can slow a robot down
A single camera frame can show where an object is, but not how it is moving or how far an interaction has progressed. World-action models address this by predicting future video alongside actions in a closed loop, which makes recent visual history a natural input for control. The catch is that processing more history adds computation, and that delay can undermine the very responsiveness the history was meant to improve.
The paper therefore asks two linked questions: how does control performance scale with the length of visual context, and can those gains survive real-time constraints? Long-WAM is presented as a model-system framework rather than a single network, because the answer requires both a pretraining recipe that makes history useful and an execution stack that keeps latency bounded.
Autoregressive pretraining makes history usable
The paper's key distinction is between accessing history and learning to predict from it. An autoregressive (AR) video model is trained to predict each next chunk from preceding chunks, so it learns a causal, history-to-future factorization. Long-WAM first learns this causal prediction from robot and egocentric videos without action labels, then adapts to actions while preserving that structure.
The comparison that carries the argument uses three video initializations on RoboCasa GR-1. Both AR-pretrained foundations gain from extending context from 0.0 to 19.2 seconds, whereas a bidirectionally pretrained initialization does not. The robot-domain AR model's lead over the bidirectional one grows from 3.3 points with no history to 17.1 points at 19.2 seconds. The authors interpret this as evidence that AR pretraining supplies a basis for using history, not merely storing it; that reading is their explanation, while the gap itself is the measured result.
Gains appear on GR-1 and LIBERO-Long
On RoboCasa GR-1 Tabletop, moving the history window from 0.0 to 19.2 seconds raises success from 63.3% to 78.7%, a 15.4-point improvement. On LIBERO-Long, 2.4 seconds of history improves success from 94.5% to 99.5%. These are different benchmarks with different task mixes and different context lengths, so the two numbers should not be read as one scale of benefit.
The paper also reports the best results among compared methods on LIBERO-Long, RoboTwin 2.0 (94.4% average success), and DOMINO (34.9% success). DOMINO adds moving objects, which is where recent history matters most for inferring motion. The pretraining corpus is described as roughly 10,000 window-equivalent hours, an estimate that assigns 16 seconds to each of 2,294,889 training samples rather than a count of unique raw footage.
Longer context costs latency, but not linearly in usefulness
Long context is only deployable if the controller still responds in time, so the paper measures end-to-end inference latency on an RTX 5090. With four denoising steps per expert, each action chunk including future-video latent prediction takes 107.4 ms at 2.4 seconds of history, a 3.3× speedup over BF16 eager execution. The same pipeline takes 74.6 ms with no history and 341.0 ms at 19.2 seconds, so an eightfold increase in history raises latency by about 3.2×.
Three system choices support this. Streaming observation encoding runs the VAE on incoming frames while the robot is still executing, moving prefix work ahead of the inference trigger. Asynchronous execution overlaps inference with ongoing motion. Hardware-specific acceleration, including NVFP4 quantization and device-specific kernel tuning, targets RTX 5090, DGX Spark, and Jetson AGX Thor. The paper states these adaptations reflect hardware constraints rather than device-exclusive algorithms, and reports total speedups of 3.2–4.1× over BF16 eager execution.
Dynamic manipulation separates Long-WAM from baselines
The real-robot evaluation is where the context argument becomes concrete. On a Unitree G1, Long-WAM grasps cups from a conveyor moving at 7.5 cm/s in 90% of trials and stacks moving cups in 95%. In the dynamic cup-stacking setting, the paper reports that Pi0.5 and Fast-WAM succeed in none of 20 trials. On YAM, Long-WAM completes tasks lasting over 40 seconds on average with 81.7% success, extending the evaluation from fast interception to sustained execution.
These results should be read as task-specific comparisons rather than a general ranking of the baselines. The paper's own framing is that a low-level controller benefits from remembering motion and recent interaction state, while hour-scale goals, reasoning, and task decomposition belong to a higher-level planner. Long-WAM is positioned as a memory-informed executor that complements such planning in composite tasks.
The useful window depends on data and compute
The paper does not claim that longer is always better. Performance declines at 38.4 seconds, and the authors attribute this to sparse real history in current datasets, reporting 80.4% padding at that length. That explanation points to denser and longer robot recordings as the route to extending the useful window, rather than to a fixed memory optimum.
A second boundary is methodological: each context window in the study is trained as its own model, which gives a clean view of how much history helps at every length but does not demonstrate a single policy that adjusts its window at test time. The measured latency profile, from 74.6 ms without history to 341.0 ms at 19.2 seconds, indicates how much computation such adaptivity could reclaim. The authors also note that the training surrogate for future latents retains a residual ground-truth component, which they describe as a training approximation rather than identical conditioning at inference.
Design context, prediction, and latency together
The paper's most decision-relevant claim is a comparison, not a single score: the payoff from longer visual history depends on how the video foundation was pretrained. On GR-1, the robot-domain AR model's advantage over a bidirectional initialization widens from 3.3 points at zero history to 17.1 points at 19.2 seconds, which means the choice of pretraining objective changes how much a longer context window is worth.
That reframes the design question from whether a policy has memory to how its context length, predictive representation, and execution budget are chosen together. The appropriate window is presented as task- and compute-dependent rather than universal, and the paper's own next step is a single policy that adjusts its window at test time. For readers comparing world-action models, the practical implication is to evaluate them jointly on what they remember, how well they act, and how quickly they incorporate new observations.
Sources used in this answer
Long-WAM: Scaling the Context of World-Action Models
Longer visual history improves robot control far more when the video foundation is pretrained autoregressively: on RoboCasa GR-1, context from 0.0 to 19.2 seconds raises success from 63.3% to 78.7%, while a bidirectionally pretrained initialization shows no net gain.