Eight robot-policy leaderboards, six different number ones
π0.5 tops three boards, all run on a fixed set of open models. The two live real-robot arenas are led by models that appear on no other top list.
Key number6 of 8Distinct #1 models across the eight policy leaderboards we track
We track eight public leaderboards for robot policies (vision-language-action models that output motor commands). Across them, six different models hold first place.
| Leaderboard | Setting | Run by | #1 | Score | Snapshot |
|---|---|---|---|---|---|
| RoboArena | Real robot | 7 universities, DROID platform | Spirit v1.6 | 1,788 (rating) | 2026-10-28 |
| RoboChallenge Table30 | Real robot | Dexmal + Hugging Face | LoopWAM | 43.33% | 2026-10-28 |
| SimFoundry | Real robot | Paper authors | π0.5 | 73.1% | 2026-09-18 |
| FlagEval-Robo | Sim + real | BAAI | π0.5 | 66.88% | 2026-08 |
| ArmnetBench v0.1 | Benchmark paper | Paper authors | π0.5 | 47.6% | 2026-08-29 |
| RoboLab-120 | Simulation | NVIDIA SRL | Cosmos3 Nano Policy | 36.8% | 2026-08-12 |
| BEHAVIOR Challenge 2025 | Simulation | Stanford SVL | Robot Learning Collective | 0.26 (Q-score) | 2025 final |
| LIBERO | Simulation | Self-reported in papers | SimpleVLA-RL (OFT) | 99.1% | 2026-09-23 |
Three patterns
- π0.5 leads where the field is fixed. Its three wins come from boards where the organisers ran a chosen set of open models themselves. It also appears in the top eight on RoboArena (as π0.5-DROID, 4th) and RoboChallenge (as a baseline, 8th), but does not lead either.
- The open, live real-robot arenas are led by entrants that appear nowhere else. Spirit v1.6 and LoopWAM do not appear in any other board's top list. Open submission brings in models that fixed-field benchmarks never test.
- Scores do not carry across boards. π0.5 scores 73.1% on SimFoundry and 14.33% as a baseline on RoboChallenge. The tasks, robots and scoring rules differ, so the numbers measure different things.
Most models are on one board only
Of 1,224 policies with published simulation scores, 85% appear on a single benchmark. Only π0.5 has a score on all eight boards above. That is why we do not average these leaderboards into one ranking: a combined score would mostly reflect which boards each model happened to enter.
The full lists, with sources, are on our robot policy page.