Rotate back to portrait for the best experience.

02 — Data

AI Model Performance

Updated 30 October 2026
i

Every value checked against its primary source on 23 Oct 2026; hover a number for its source, or select it to open the source at the figure. Paired sim/real results: SIMPLER arXiv 2405.05941 · AutoMate arXiv 2407.08028 · arXiv 2603.18532 · IndustrialVLA-Bench arXiv 2609.25562 · arXiv 2608.26821 · SO-101 benchmark arXiv 2606.08881 · ALOHA benchmark arXiv 2511.11298 · OpenVLA-OFT arXiv 2502.19645 · Cosmos Policy (ICLR 2026). Company marks are site favicons, used to identify each model. Draft for discussion.

Submit your results
Data
PDF report 7 pagesCSV spreadsheetsJSON codeMCP for AI agentsHow to cite →
?~ read from a chart

Calibrated simulators land within about 5–10 points of reality and rank policies almost identically (SIMPLER r = 0.92, PolaRiS r = 0.90). An off-the-shelf benchmark score can sit 30–40 points above what the same model does on a real arm.

Real robot, method publishedClaim, method not publishedSimulationNot reported (owed)

What changed

  1. Ten more variants from seven papers (LOGIV, R2A, QbD-VLA, VLA-Stem, JAM, behaviour context, movement-primitive chunking), built on π0.5, π0, GR00T N1.7, OpenVLA-OFT, Fast-WAM and ACT. Each checked against its paper’s own tables.
  2. New: third-party leaderboards. Ten groups that run robot models themselves, real robot and simulation, with where our scorecard models rank on each.
  3. Model pages now list variants: published work built on a model, with its results next to the base model measured in the same paper. First six papers added, built on π0, π0.5, SmolVLA, GR00T N1.7 and OpenVLA-OFT.
  4. First published: 15 robot foundation models across six axes, every value checked against its primary source.

Email me the report

The full AI Model Performance report as a PDF: every model, every criterion, sources and method.

Sent once to this address. The box above is optional. Privacy

Help keep this honest

Results, corrections, test requests. Nothing changes until it's checked against its source.

This is about

MethodologyAxes, evidence grades, sources, downloads and how to cite

Get notified

New results on AI Model Performance, and news from B2A. Nothing else.