A shorter route to a predicted future
A preprint first posted on August 12 introduces RIFT, short for Rollout-free Imagination via Future Tokens. World action models normally generate a sequence of future video states, then use those predicted states to choose robot actions. RIFT instead creates the internal future representation in one pass, so it keeps the model's future-conditioned action path without repeatedly generating video frames. The authors report that removing or scrambling that stored future information sharply reduced simulated task success, then use the finding to motivate the new one-pass method.[1]
The reported simulation result is strong on its stated terms. On all 40 LIBERO tasks, the paper reports 98.8 percent success for RIFT, close to the 98.4 to 98.6 percent reported for three rollout-based alternatives. Its latency table reports that RIFT lowers per-action-chunk latency by 68.2 to 89.1 percent versus those rollout systems, while running at 1.1 times the latency of a current-only model. The authors also report 92.9 and 92.6 percent success on RoboTwin 2.0's clean and randomized simulated settings.[1,3]
The scorecard is still a simulation result
That distinction changes what the latency result can support. LIBERO's own documentation describes 130 tasks across four research suites, with human teleoperation demonstrations and an off-screen simulated environment. RoboTwin's official repository likewise describes a bimanual manipulation platform, data and policy evaluation stack rather than a physical field trial. RIFT's authors make the boundary explicit in their limitations section: evaluation is simulation-only, and physical robots and non-world-action-model fusion backbones are future work.[1,2,3]
There is a second reason to keep the conclusion narrow. An independent audit from Toyota Technological Institute at Chicago, Argonne National Laboratory and the University of Chicago tested LIBERO, RoboTwin 2.0 and other manipulation benchmarks for shortcut solvability, statistical significance, overfitting and data-source dependence. It found a 0.09-billion-parameter probe without a language encoder scored at or near reported state of the art on LIBERO, while RoboTwin 2.0 failed fewer of its tests. That does not invalidate RIFT's result, but it means the 98.8 percent number is evidence of a faster method in this benchmark setup, not evidence that a hardware robot now has a faster, reliable imagination loop.[1,4]
The decision change is therefore technical, not commercial: teams building world-action-model research stacks have a credible lead that future conditioning may not require iterative video rollout at inference time. They cannot yet convert the reported latency saving into a deployment claim, staffing plan or safety argument. The next meaningful checkpoint is a public implementation and a physical scorecard: repeated tasks on real hardware, action latency measured end to end, failures and resets counted, and tests against new objects, viewpoints and embodiments.[1,2,3,4]