← Leaderboard
π0.5
VLAPhysical Intelligence · Apr 22, 2025
Successor to π0 trained for open-world generalization, with open weights through openpi. It is the most widely used common baseline across 2026 leaderboards.
- Type
- VLA
- Parameters
- 3.3B
- Backbone
- PaliGemma 3B
- Weights
- Open
- License
- Gemma
- Commercial use
- Unclear
- Released
- Apr 22, 2025
- Robots
- DROID Franka, SO-101
- SO-101
- Yes
- LeRobot
- Supported
Benchmark results
12 of 13 benchmarksRoboLab-120Simulation
- Official leaderboard ↗Nov 17, 2026· n = 1200
#15 of 2228.0%
RoboCasa365Simulation
- Official leaderboard ↗Apr 2, 2026
#24 of 2616.9%
RoboDojo (sim)Simulation
- Paper / model card ↗Jul 2026
#29 of 406.9%, self-reported by the authors, not a leaderboard entry
LIBERO-PlusSimulation
- Paper / model card ↗Jun 2026
#10 of 2485.6%, self-reported by the authors, not a leaderboard entry
LIBERO-PROSimulation
- Paper / model card ↗Oct 2025
#2 of 662.0%, self-reported by the authors, not a leaderboard entry
- Official leaderboard ↗Aug 10, 2026· n = 10000
#17 of 2358.3%
- Paper / model card ↗Jul 2026· ABC→D
#5 of 114.69 / 5, self-reported by the authors, not a leaderboard entry
- Paper / model card ↗Sep 3, 2025
#35 of 4696.8%, self-reported by the authors, not a leaderboard entry
RoboArenaReal robot
- Official leaderboard ↗Nov 17, 2026· n = 745
#4 of 131608
RoboChallenge Table30 V2Real robot
- Official leaderboard ↗Apr 15, 2026· V2· unverified
#8 of 1214.3%, second-hand number, not confirmed at the source
RoboDojo-RealEvalReal robot
- Paper / model card ↗Jul 2026· n = 180
#5 of 1212.8%, self-reported by the authors, not a leaderboard entry
SO-101 BenchLow-cost arm
- Paper / model card ↗Oct 27, 2026· n = 916
#7 of 1141.0%, self-reported by the authors, not a leaderboard entry
Not reported on 1 benchmark
Robustness drop: 96.8% on LIBERO → 85.6% on LIBERO-Plus → 62.0% on LIBERO-PRO. The perturbed variants show how much of the LIBERO score is memorization.
Adoption
updated Nov 17, 2026Speed per task
| Setting | Per task | Per planner call | Calls | Cost | Source |
|---|---|---|---|---|---|
| SO-101 real robot, 45 Sim2Real layouts | 59 s | – | – | – | Oct 27, 2026 ↗ |
| Real Franka (DROID setup), 12 tasks × 9 trials | 62 s | – | – | – | Nov 2026 ↗ |
| RoboDojo-RealEval, 18 tasks × 10 trials · RTX 4090 policy server on LAN | 67 s* | – | – | – | Jul 2026 ↗ |
Running the model
| Hardware | Precision | Config | Latency / chunk | Steps | VRAM | Chunk | Control | Source |
|---|---|---|---|---|---|---|---|---|
| B200 | bf16 | 10 steps, guidance off (DROID) | 31.99 ms | 10 | – | 15 | 15 Hz | Sep 2026 ↗ |
| H200 | bf16 | 10 steps, guidance off (DROID) | 46.96 ms | 10 | – | 15 | 15 Hz | Sep 2026 ↗ |
| L40S | – | RoboLab-120 evaluation (DROID) | – | – | – | 15 | – | May 9, 2026 ↗ |
| RTX 4090 | bf16 | vanilla (synchronous) inference | 76 ms | 5 | – | 50 | 50 Hz | Jun 2025 ↗ |
| RTX 5090 | bf16 | 10 steps, guidance off (DROID) | 34.92 ms | 10 | – | 15 | 15 Hz | Sep 2026 ↗ |
| RTX PRO 6000 Blackwell | bf16 | 10 steps, guidance off (DROID) | 57.7 ms | 10 | – | 15 | 15 Hz | Sep 2026 ↗ |
Trained on
Wrong or missing number? Report an error