π0
Open weightsPhysical Intelligence
2 of 6 axes published · values checked against their primary sources on 23 Oct 2026
Success
Does it do the trained task? (% of trials, trained setup)
72.3%
Clean dish 87%, sponge in pot 83%, unzip bag 67%, fold shorts 52% (200 demos); 50 trials per task.
Source: Benchmarking VLAs on ALOHA, arXiv 2511.11298 · Table 3, macro-average
Robustness
Does it survive the cell drifting? (% under the perturbation set)
58.5%
Unseen objects at random positions: 81%, 73%, 37%, 43% — 13.8 points below in-distribution. Highest robustness of any model with real trial counts.
Source: Benchmarking VLAs on ALOHA, arXiv 2511.11298 · Table 5 (mean of 4 tasks)
Instruction
Does rewording the instruction break it? (% under paraphrase)
Owed
Not reported.
Recovery
Does it fix its own mistakes? (% of injected failures recovered)
Owed
Not reported.
Speed
Is it as fast as a trained person? (human cycle time ÷ robot cycle time)
Owed
No human comparison. Time to success 13–85 s across successful trials.
Interventions
How often does a person step in? (per 1,000 cycles)
Owed
Not reported.
Built on π0
Published work that starts from π0: fine-tunes, add-ons and models trained from it. Their results are listed here, apart from π0's own values above, with the source each one comes from.
QbD-VLA
Inference settingOpen codeColorado School of Mines · Northeastern University · USC Institute for Creative Technologies · University of Central Florida · 28 Oct 2026
Shrinks the model by storing most of its numbers at low precision, choosing which parts to compress by testing the robot’s task success in simulation.
- Success73.5%−1.1 pts vs π0 in the same paper (74.6%)Simulation · LIBERO, 4 suites, ~4-bit weights · 800 trials · Grade BBase: the LeRobot LIBERO fine-tune at full precision, which scores well below π0’s own published LIBERO results.
Source: QbD-VLA paper, arXiv 2610.38661
Backtrack
Add-onStanford · ETH Zurich · 23 Oct 2026
A small add-on spots a failing attempt, reverses the arm's last moves and lets the unchanged model try again. Tested on π0-FAST.
- Success93.0%+1.7 pts vs π0 in the same paper (91.3%)Simulation · LIBERO, 4 suites · Grade B
- Recovery19.6%Simulation · LIBERO, 4 suites · Grade BShare of the base model's failures turned into successes.
Horizon–depth coupling
Inference settingCASIA · FiveAges · NUS · 23 Oct 2026
No retraining: when the arm runs longer before re-planning, the model takes fewer sampling steps per move.
- Success95.5%+3.1 pts vs π0 in the same paper (92.4%)Simulation · LIBERO average · Grade B500 episodes per suite. On RoboTwin 2.0: 45.7% vs 37.0%.
Source: Execution horizon and sampling depth, arXiv 2610.32102
Know of one we missed? Send it with its source.