Rotate back to portrait for the best experience.

02 — Data · AI Model Performance

π0.5

Open weightsPhysical Intelligence

4 of 6 axes published · values checked against their primary sources on 23 Oct 2026

  1. Success

    Does it do the trained task? (% of trials, trained setup)

    56.3%

    Real robotGrade B · Paper or report with method (rubric or error bars)80 trialsSO-101 · 4 tasks

    Average of pen transfer 95%, colour sorting 10%, packing 55%, pen placement 65%; 100 demos per task. π0.5's own paper reports 60–80% task progress in unseen homes (Fig. 7b, 10 trials per task per home).

    Source: SO-101 VLA benchmark, arXiv 2606.08881 · Results table, average

  2. Robustness

    Does it survive the cell drifting? (% under the perturbation set)

    85.3%

    SimulationGrade B · Paper or report with method (rubric or error bars)LIBERO-Plus

    Non-language perturbations (camera, robot state, lighting, layout), simulation.

    Source: IndustrialVLA-Bench, arXiv 2609.25562 · Table 3

  3. Instruction

    Does rewording the instruction break it? (% under paraphrase)

    71.3%

    SimulationGrade B · Paper or report with method (rubric or error bars)LIBERO-Para

    Paraphrased instructions, 25.2 points below clean. Real-world: π0.5 paper Fig. 9 shows ~85% language following on familiar objects, ~70% on new ones (read from chart).

    Source: IndustrialVLA-Bench, arXiv 2609.25562 · Table 4

  4. Recovery

    Does it fix its own mistakes? (% of injected failures recovered)

    30.8%

    Real robotGrade B · Paper or report with method (rubric or error bars)SO-101

    Share of failures the policy recovered from without help.

    Source: SO-101 VLA benchmark, arXiv 2606.08881 · Recovery results

  5. Speed

    Is it as fast as a trained person? (human cycle time ÷ robot cycle time)

    Owed

    No cycle time against a person. Related, not equivalent: 58.8 ms per policy call (Table 5). Home tasks take 2–5 minutes, no human comparison.

    Checked: IndustrialVLA-Bench, arXiv 2609.25562

  6. Interventions

    How often does a person step in? (per 1,000 cycles)

    Owed

    Not reported. The π0.5 paper scores task progress, not interventions.

    Checked: π0.5 paper, arXiv 2504.16054

Built on π0.5

Published work that starts from π0.5: fine-tunes, add-ons and models trained from it. Their results are listed here, apart from π0.5's own values above, with the source each one comes from.

  1. LOGIV

    Add-onOpen code

    CUHK-Shenzhen · A*STAR · Caltech · NUS · HKUST · COCO Matrix · NTU · 29 Oct 2026

    A planning layer above the model keeps track of which steps depend on which, checks them as the robot works, and repairs only the broken part of the plan.

    • Success56.7%+16.7 pts vs π0.5 in the same paper (40%)Real robot · 3 two-arm tasks, Astribot S1 and AgileX Piper · 30 trials · Grade B17/30 vs 12/30, same π0.5 checkpoint. In simulation (RoboTwin 2.0, no disturbances): 72% vs 56% (Table I).
    • Robustness60%+28 pts vs π0.5 in the same paper (32%)Simulation · RoboTwin 2.0, injected disturbances · 100 trials · Grade B10 tasks × 10 initial states. On BiCoord: 48% vs 31%.
    • Recovery72%+54 pts vs π0.5 in the same paper (18%)Simulation · RoboTwin 2.0, restoring a broken step · Grade BShare of recoverable disturbances that were put right. On BiCoord: 68% vs 21%.

    Source: LOGIV paper, arXiv 2610.40016

  2. Behavior context

    Fine-tuning recipe

    Stanford University · 29 Oct 2026

    Shows the model a few examples of how the arm responds under the current controller settings, so one model adapts to different controllers.

    • Robustness50.3%+32.8 pts vs π0.5 in the same paper (17.5%)Real robot · WidowX, controller types not seen in training · Grade BBase: π0.5 fine-tuned on the same data without the examples. 60 trials per held-out controller type. Unseen gain settings: 52.0% vs 28.5%.

    Source: Cross-controller paper, arXiv 2610.39280

  3. QbD-VLA

    Inference settingOpen code

    Colorado 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.

    • Success96.2%−0.6 pts vs π0.5 in the same paper (96.8%)Simulation · LIBERO, 4 suites, ~4-bit weights · 800 trials · Grade BBase: the LeRobot LIBERO fine-tune at full precision. On a real Franka arm (Table 5) the compressed model succeeded 0 of 6 times, against 6 of 6 at full precision.

    Source: QbD-VLA paper, arXiv 2610.38661

  4. VLA-Stem

    Add-on

    USTC · University of Macau · Peking University · 27 Oct 2026

    A small safety layer follows the model’s planned motion and adjusts it on the fly to avoid fixed and moving obstacles.

    • Success73%+3.8 pts vs π0.5 in the same paper (69.2%)Simulation · ManiSkill, 3 tasks, no obstacles · 9,600 trials · Grade BMean of PickCube, Object Stowing and StackCube (our average). 3,200 episodes per task.
    • Robustness52.8%+46.7 pts vs π0.5 in the same paper (6.1%)Simulation · Same 3 tasks, fixed and moving obstacles · 19,200 trials · Grade BSuccess without a collision, mean of 6 obstacle settings (our average). Real-robot results are shown only as ranges (Fig. 5).

    Source: VLA-Stem paper, arXiv 2610.37378

  5. R2A

    Add-onOpen code

    VinUniversity · VinRobotics · DFKI · TU Darmstadt · 27 Oct 2026

    Looks up similar moments in past successful demonstrations and blends those motions into the frozen model’s actions to keep it on track when conditions change.

    • Robustness90.4%+4.5 pts vs π0.5 in the same paper (85.9%)Simulation · LIBERO-Plus, 7 perturbation types · Grade Bopenpi’s pi05_libero checkpoint. On two-arm RoboTwin 2.0-Plus: 66.2% vs 56.8% (Table 2).
    • Instruction86.6%+0.8 pts vs π0.5 in the same paper (85.8%)Simulation · LIBERO-Plus, changed instructions · Grade B

    Source: R2A paper, arXiv 2610.37215

  6. JevVLA

    Add-on

    Zhejiang University · Cornell · DEEP Robotics · NUS · 26 Oct 2026

    A tiny add-on (about 80,000 parameters) ranks and nudges the frozen model's proposed actions.

    • Success77.1%+1.3 pts vs π0.5 in the same paper (75.8%)Simulation · MetaWorld MT50 · 2,500 trials · Grade B
    • Robustness80.2%+1.5 pts vs π0.5 in the same paper (78.6%)Simulation · LIBERO-Plus, all perturbations · Grade BCamera, lighting, layout, start pose, sensor noise, background and language changes.
    • Instruction81.3%+1.6 pts vs π0.5 in the same paper (79.7%)Simulation · LIBERO-Plus, language changes · 1,537 trials · Grade BChanged instructions, not strictly paraphrases.

    Source: JevVLA paper, arXiv 2610.35010

  7. RoboDuo

    Add-on

    Robotera · HUST · Tsinghua · 25 Oct 2026

    A small coding agent checks whether the task got done and, if not, hands the missing steps to a second π0.5.

    • Success98.4%+1.2 pts vs π0.5 in the same paper (97.3%)Simulation · LIBERO average · Grade B
    • Robustness39.3%+7.3 pts vs π0.5 in the same paper (32.0%)Simulation · LIBERO-PRO, 8 conditions · Grade B500 episodes per condition.
    • Recovery41%Simulation · LIBERO-Long · Grade B'Removes 41% of the failures of π0.5.'

    Source: RoboDuo paper, arXiv 2610.34108

  8. Backtrack

    Add-on

    Stanford · 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.

    • Success97.8%+0.8 pts vs π0.5 in the same paper (97.0%)Simulation · LIBERO, 4 suites · Grade B5 seeds × 50 episodes per task.
    • Recovery26.2%Simulation · LIBERO, 4 suites · Grade BShare of the base model's failures turned into successes.

    Source: Backtrack paper, arXiv 2610.32941

  9. Horizon–depth coupling

    Inference setting

    CASIA · FiveAges · NUS · 23 Oct 2026

    No retraining: when the arm runs longer before re-planning, the model takes fewer sampling steps per move.

    • Success98%+1.1 pts vs π0.5 in the same paper (96.9%)Simulation · LIBERO average · Grade B500 episodes per suite. On RoboTwin 2.0 it is lower: 70.8% vs 73.8%.

    Source: Execution horizon and sampling depth, arXiv 2610.32102

  10. fp32 fine-tuning fix

    Fine-tuning recipe

    UT Austin · 22 Oct 2026

    Keeps a full-precision copy of the weights while fine-tuning, so small updates are no longer rounded away.

    • Success97.7%+8.7 pts vs π0.5 in the same paper (89%)Simulation · LIBERO-Object, fine-tuned · 500 trials · Grade BBase = fine-tuned with LeRobot's default recipe. 2 seeds.

    Source: fp32 master copy paper, arXiv 2610.31859

Know of one we missed? Send it with its source.

Compare in the scorecard →Submit a result or correctionHow values are graded

Get notified

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