PACT-WAM is a new robot control model that plans its next 16 actions and simulates what the world will look like as a result, without the usual computational overhead.
The system, described in a paper posted to arXiv, compresses a robot's recent visual history using a hierarchical scheme: older frames get coarse representations, recent ones get finer detail, adding up to just 256 tokens per camera view across 16 retained observations - 75% fewer tokens than standard dense encoding. A shared flow-based module then generates a 16-step action sequence and a matching visual forecast together, using separate output heads for actions and images under causal attention. A decoder reconstructs the predicted future frames from multiple camera angles. On its own, PACT-WAM hit average success rates of 98.6% on the LIBERO benchmark, 92.3% on RoboTwin 2.0, and 78.0% on real-world tasks with a Piper robot arm. Adding an optional review step, where a vision-language model screens the forecasts and rejects bad plans before execution, pushed those numbers to 99.5%, 93.4%, and 86.7%.
Robots that preview the consequences of an action before taking it tend to make fewer mistakes, but storing and processing that visual foresight is expensive - which is why many manipulation systems skip it or pay for it with slower inference. PACT-WAM's bet is that compressing history unevenly, more detail for recent frames and less for older ones, keeps most of the benefit of foresight at a quarter of the token cost. That is the kind of efficiency gain that determines whether foresight-based planning ever runs on a robot's own hardware instead of a data center.
The real-world number is the one to watch: 78% success climbing to 86.7% with review is solid, but it still trails the near-perfect scores the model racks up in simulation, a reminder that the gap between a benchmark and a warehouse floor has not closed yet.