QuantWAMs cuts robot-model memory to 29%

QuantWAMs reduced targeted world-action model blocks to about 29% of FP16 peak weight-and-activation memory while producing 1.4–1.6× block-level speedups. The research preprint says its mostly four-bit configuration stayed within 0.2–0.7 percentage points of FP16 simulation averages and ran three manipulation tasks on an AgiBot G2, but it does not report an end-to-end robot speedup. Robotics teams building simulation-first hardware testing funnels should treat quantization as a deployment lever worth reproducing on their own closed-loop workloads, not infer system-wide latency from isolated block gains.