Fast-WAM

World-action model from Tsinghua IIIS and Galaxea AI that co-trains on video during training but skips future-frame generation at inference, predicting actions in a single forward pass. It is the base of several later WAMs (CSWAM, DiffWAM, RelayWAM, Elastic-WAM). The RoboTwin and RoboDojo teams run it as a baseline on their boards.

Tsinghua University & Galaxea AI · Mar 17, 2026 · World-actionReport an error
97.6%LIBERO · #44 of 83
39.8%RoboTwin 2 · #24 of 33
2.0%Dojo Sim · #53 of 71
6BParameters

Benchmark results3

#53 of 712.0%
RoboTwin 2.0Near saturationSimulation
#24 of 3339.8%
LIBEROSaturatedSimulation
#44 of 8397.6%, self-reported by the authors, not a leaderboard entry
Not reported on 12 benchmarks
Wrong or missing number? Report an error