SWAM

World-action model from Shanghai Jiao Tong University, Shanghai Innovation Institute, Noematrix, Westlake and Tongji, pretrained on about 47,000 hours of data, that keeps the video model’s language–video alignment while adding action supervision. The paper says it releases data, training code and weights. The training code is public, and so is part of the pretraining data (Ego4D and EgoSuite are excluded for licensing, and the UMI data is still to be released). No weights link was found as of 2026-11-18.

Shanghai Jiao Tong University et al. · Nov 9, 2026 · World-actionReport an error
1Benchmarks reported
Nov 9, 2026Released
UnknownWeights
World-actionType

Benchmark results1

RoboTwin 2.0Near saturationSimulation
  • 74.8%, self-reported by the authors, not a leaderboard entryPaper / model card ↗Nov 2026· self-reported, clean-only training· n = 10000· non-standard
74.8%, self-reported by the authors, not a leaderboard entrynot ranked
Not reported on 14 benchmarks
Wrong or missing number? Report an error