← Robotics board · written 2026-10-14 by claude-opus-5-5 from the pipeline's evidence only

Yeonseok Lee (ylee-sling): Robotics · Skoor 55 · n=3 · confidence 0.7

What they built

Lee owns the official PyTorch reference implementation of Vision-Predictor-Action (VPA). VPA is a goal-conditioned robot control loop that plans in a learned latent space, picks a symbolic skill and generates an action chunk with a few flow-matching steps [source]. The README says one decision step needs K + 3 sequential network evaluations, independent of chunk length H [source]. The code is released under GPL-3.0 [source]. The README labels it "a research prototype with equation-level verificati…", and the excerpt cuts off there [source].

Record

Measured signals (n=3):
  • research.code_author (weight 20): the repo is the code for the VPA paper (2026-10) [source].
  • bio.founder_language (weight 20): the bio reads "CEO at SLING AI Inc. | PhD in Informatics" [source].
  • research.recent_first_author (weight 15): Lee is first author on the VPA paper [source].

    • Other measured facts:

  • The repo has 0 stars, 0 forks and 27 commits in 90 days [source].
  • The profile shows 81 commits this year, 0 followers and an account created 2021-03-26 [source].
  • The paper was published 2026-10-11 with no affiliation listed [source].

    • Inference: the paper is listed at author positions 0 and -1, which may mean Lee is the sole author [source].

    Inference: the repo was created 2026-10-08 and the paper went up three days later, so the public record is about one week deep [source].

    Why now

    The README argues that VLA models emit actions token by token through a large language model, so the number of forward passes grows with sequence length and slows robot reaction [source]. VPA removes that autoregressive decoder [source]. Market or customer timing is not in the evidence.

    Unknowns

  • What SLING AI Inc. sells is not in the evidence. The profile company field is empty [source].
  • Benchmark or physical-robot results are not in the evidence.
  • Funding, team size, customers and PhD institution are not in the evidence.
  • Inference: a GPL-3.0 license on the core method may limit closed commercial use, and the plan for this is unclear [source].

    • Questions for the room

  • The README claims K + 3 evaluations regardless of H. What wall-clock latency and task success did you measure on a physical robot against a VLA baseline?
  • Your bio says CEO at SLING AI Inc. What does the company sell, who pays for it, and how does a GPL-3.0 release fit that plan?
  • The paper lists no affiliation, and you may be the only author. Who else built VPA, and where did you earn the PhD?

    • What would move the Skoor

    Raise: hardware latency and success numbers, or named co-founders or customers at SLING AI. Lower: no experiments beyond equation-level checks, or no operating company behind the CEO title.