Real world
The real-robot test
Simulation tests are not enough: a robot has to prove itself with real objects. On RoboArena, researchers from several labs give the same task to two identical robots (Franka armFrankaA German robotic arm widely used in research labs., DROIDDROIDA large dataset of robot videos recorded in dozens of labs, and the standard platform (Franka arm + cameras) that goes with it. platform), each controlled by a different model, without knowing which is which, then pick the winner (blind A/B comparisonBlind A/B comparisonAn evaluator watches two robots perform the same task without knowing which is which, then picks the better one.).
The results produce an Elo ratingElo ratingA head-to-head score, as in chess: you gain points by beating others. A 100-point gap ≈ 64% chance of winning., as in chess: it shows who beats whom, not a success percentage. Rule of thumb: a 100-point gap ≈ a 64% chance of winning a head-to-head matchup.
Who beats whom?
Official leaderboard: policies with at least 100 A/B evaluations, as on the RoboArena site
| # | Policy | Rating (Elo) | Range ± 1 standard deviation | Matchups won vs. #1* | A/B evaluations | Availability |
|---|---|---|---|---|---|---|
| 1 | DreamZero UnspecifiedOpen-source | 1735 ± 43 | reference | 190 | ○ Offline 0% over 30 days | |
| 2 | π0.5 (version DROID) Physical Intelligence | 1608 ± 31 | 32% | 745 | ● Online 100% over 30 days | |
| 3 | π0-FAST (version DROID) Physical Intelligence | 1582 ± 30 | 29% | 941 | ● Online 100% over 30 days | |
| 4 | PaliGemma · vq (DROID) UnspecifiedOpen-source | 1546 ± 30 | 25% | 952 | ● Online 100% over 30 days | |
| 5 | PaliGemma · diffusion (DROID) UnspecifiedOpen-source | 1533 ± 30 | 24% | 949 | ● Online 100% over 30 days | |
| 6 | PaliGemma · fast specialist (DROID) UnspecifiedOpen-source | 1531 ± 30 | 24% | 1,057 | ○ Offline 0% over 30 days | |
| 7 | PaliGemma · fast (DROID) UnspecifiedOpen-source | 1516 ± 30 | 22% | 1,062 | ○ Offline 0% over 30 days | |
| 8 | π0 (version DROID) Physical Intelligence | 1461 ± 30 | 17% | 1,120 | ○ Offline 0% over 30 days | |
| 9 | PaliGemma · binning (DROID) UnspecifiedOpen-source | 786 ± 50 | 0% | 629 | ○ Offline 0% over 30 days |
* Estimate derived from the ratings: out of 100 matchups against DreamZero (official #1), how many the policy would win. 50% = equal level.
Are the rankings shifting?
Week-by-week Elo rating of the top 5 on the official leaderboard
Show chart data: Elo rating of the top 5 official policies, by date
| Date | DreamZero | π0.5 (version DROID) | π0-FAST (version DROID) | PaliGemma · vq (DROID) | PaliGemma · diffusion (DROID) |
|---|---|---|---|---|---|
| Jun 8 | 1739 Elo | 1612 Elo | 1580 Elo | 1551 Elo | 1540 Elo |
| Jun 15 | 1739 Elo | 1612 Elo | 1579 Elo | 1552 Elo | 1540 Elo |
| Jun 22 | 1739 Elo | 1615 Elo | 1580 Elo | 1552 Elo | 1537 Elo |
| Jun 29 | 1733 Elo | 1611 Elo | 1573 Elo | 1545 Elo | 1528 Elo |
| Jul 6 | 1737 Elo | 1615 Elo | 1578 Elo | 1549 Elo | 1532 Elo |
| Jul 13 | 1737 Elo | 1613 Elo | 1580 Elo | 1548 Elo | 1533 Elo |
| Jul 20 | 1736 Elo | 1611 Elo | 1581 Elo | 1548 Elo | 1533 Elo |
| Jul 27 | 1720 Elo | 1594 Elo | 1566 Elo | 1533 Elo | 1515 Elo |
| Aug 3 | 1721 Elo | 1593 Elo | 1567 Elo | 1534 Elo | 1518 Elo |
| Aug 10 | 1723 Elo | 1595 Elo | 1569 Elo | 1537 Elo | 1520 Elo |
| Aug 17 | 1725 Elo | 1599 Elo | 1570 Elo | 1536 Elo | 1521 Elo |
| Aug 24 | 1969 Elo | 1843 Elo | 1815 Elo | 1780 Elo | 1765 Elo |
| Aug 31 | 1736 Elo | 1609 Elo | 1582 Elo | 1547 Elo | 1533 Elo |
| Sep 7 | 1735 Elo | 1608 Elo | 1582 Elo | 1547 Elo | 1532 Elo |
| Sep 14 | 1735 Elo | 1608 Elo | 1582 Elo | 1546 Elo | 1533 Elo |
| Sep 21 | 1735 Elo | 1608 Elo | 1582 Elo | 1546 Elo | 1533 Elo |
| Sep 25 | 1735 Elo | 1608 Elo | 1582 Elo | 1546 Elo | 1533 Elo |
Source: RoboArena · paper (arXiv 2506.18123) · leaderboard updated on Sep 25, 2026, 3:50 PM