What does Learned Motion Matching (Holden et al. 2020) replace, what data and training did it use, and what limits do the authors report?

Learned Motion Matching (2020) replaces the stored animation database and nearest-neighbour search of motion matching with three small networks (Decompressor, Stepper, Projector) so the runtime keeps no pose database; memory is the stated motive. Training runs on CPU for up to 500,000 iterations (a few hours for results, overnight for best). Reported scenarios span 22,398 to 694,272 frames, but the 694k Bear set was augmented from about one minute (3,581 frames) of key-framed animation. In the user study subjects struggled to identify the reference system, with a slight responsiveness preference for basic motion matching and no quality preference; the authors found Projector accuracy had the largest impact on quality. Curator qualifications: Paper and help pages read 2026-10-22; no network trained, no cloth exported. Exact wording nuance: the study reports a trend toward correct identification in part one; "could not reliably tell" summarises part two. Production adoption beyond the authors' AAA study not assessed. Limitations: Paper and help pages read 2026-10-22; no network trained, no cloth exported. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://theorangeduck.com/media/uploads/other_stuff/Learned_Motion_Matching.pdf",
      "note": "[producer working copy; not public] sha256 534c1993e3f2f111 (equal to producer pin S8); pdftotext lines 343-346, 550-557, 650-662, 690-720, 770-778; [producer working copy; not public] sha256 534c1993e3f2f111 (equal to producer pin S8); pdftotext lines 1250-1284, 1488-1494"
    }
  ]
}