Poke &Wiggle

Leaderboard

Model

DiT-Flow

Carnegie Mellon University · UC Berkeley

#3

Rank · PAW-GEN-10

Model information

Website
Website⁠ (opens in a new tab)
Paper
arXiv⁠ (opens in a new tab)
Code
SudeepDasari/dit-policy⁠ (opens in a new tab)
Parameters
115M
Released
October 2024
Licence
MIT

PAW-GEN-10

Share of episodes rated Ok or Excellent.

Summary

Overall Success
19%
Evaluations
3600
NominalEvaluated on the placements the model was trained on.InterpolationEvaluated on unseen placements between the training placements.ExtrapolationEvaluated on unseen placements outside the training placements' convex hull.
EnvironmentNominal · D10Trained on ~10 demonstrations per environment.Nominal · D100Trained on ~100 demonstrations per environment.Nominal · D300Trained on ~300 demonstrations per environment.Interpolation · D10Trained on ~10 demonstrations per environment.Interpolation · D100Trained on ~100 demonstrations per environment.Interpolation · D300Trained on ~300 demonstrations per environment.Extrapolation · D10Trained on ~10 demonstrations per environment.Extrapolation · D100Trained on ~100 demonstrations per environment.Extrapolation · D300Trained on ~300 demonstrations per environment.
Basic skills
Sort screws into bins7%#195% CI 2%–21%0%#495% CI 0%–11%0%#395% CI 0%–11%0%#195% CI 0%–11%0%#195% CI 0%–11%3%#295% CI 1%–17%Coming soon
Datum-corner alignment3%#395% CI 1%–17%33%#395% CI 19%–51%27%#495% CI 14%–44%0%#495% CI 0%–11%20%#395% CI 10%–37%17%#495% CI 7%–34%Coming soon
Pour screws into the funnel23%#295% CI 12%–41%60%#395% CI 42%–75%87%#295% CI 70%–95%0%#495% CI 0%–11%53%#395% CI 36%–70%67%#295% CI 49%–81%Coming soon
High precision
Connect DC jack0%#195% CI 0%–11%7%#395% CI 2%–21%0%#295% CI 0%–11%0%#195% CI 0%–11%3%#195% CI 1%–17%0%#195% CI 0%–11%Coming soon
Route cable through closed hoops0%#195% CI 0%–11%0%#395% CI 0%–11%0%#395% CI 0%–11%0%#195% CI 0%–11%0%#495% CI 0%–11%0%#395% CI 0%–11%Coming soon
Put tools in standing toolbox0%#295% CI 0%–11%3%#195% CI 1%–17%3%#395% CI 1%–17%0%#195% CI 0%–11%0%#295% CI 0%–11%0%#395% CI 0%–11%Coming soon
Bimanual
Place loaded boxes in container (bimanual)3%#295% CI 1%–17%30%#295% CI 17%–48%73%#295% CI 56%–86%0%#295% CI 0%–11%33%#195% CI 19%–51%53%#295% CI 36%–70%Coming soon
Place spray bottles upright, nozzle forward (bimanual handover)7%#495% CI 2%–21%90%#195% CI 74%–97%100%#195% CI 89%–100%0%#395% CI 0%–11%90%#195% CI 74%–97%90%#195% CI 74%–97%Coming soon
Tool use
Open the toolbox with a screwdriver0%#195% CI 0%–11%30%#295% CI 17%–48%30%#195% CI 17%–48%3%#195% CI 1%–17%7%#395% CI 2%–21%27%#295% CI 14%–44%Coming soon
Clamp electrical component0%#195% CI 0%–11%23%#295% CI 12%–41%3%#295% CI 1%–17%0%#195% CI 0%–11%0%#295% CI 0%–11%0%#295% CI 0%–11%Coming soon
Pooled4%#395% CI 3%–7%28%#295% CI 23%–33%32%#395% CI 27%–38%0%#495% CI 0%–2%21%#395% CI 16%–26%26%#395% CI 21%–31%Coming soon

Data Scaling

Scaling results for DiT-Flow: SuccessDiT-Flow nominal: D10 4% (95% CI 3%–7%), D100 28% (95% CI 23%–33%), D300 32% (95% CI 27%–38%). DiT-Flow interpolation: D10 0% (95% CI 0%–2%), D100 21% (95% CI 16%–26%), D300 26% (95% CI 21%–31%).Success rate (%)03060DiT-Flow nominal: 4% (95% CI 3%–7%) on D10DiT-Flow nominal: 28% (95% CI 23%–33%) on D100DiT-Flow nominal: 32% (95% CI 27%–38%) on D300DiT-Flow interpolation: 0% (95% CI 0%–2%) on D10DiT-Flow interpolation: 21% (95% CI 16%–26%) on D100DiT-Flow interpolation: 26% (95% CI 21%–31%) on D300D10D100D300

Spatial Generalization

Spatial generalization for DiT-Flow: SuccessDiT-Flow: Nominal 21% (95% CI 19%–24%), Interpolation 16% (95% CI 13%–18%), Extrapolation coming soon.Success rate (%)03060Coming soonDiT-Flow: 21% (95% CI 19%–24%) on NominalDiT-Flow: 16% (95% CI 13%–18%) on InterpolationNominalInterpolationExtrapolation

Per Environment (Success)

Model performance by environmentDiT-Flow overall: Sort screws into bins 2%, Datum-corner alignment 17%, Pour screws into the funnel 48%, Connect DC jack 2%, Route cable through closed hoops 0%, Put tools in standing toolbox 1%, Place loaded boxes in container (bimanual) 32%, Place spray bottles upright, nozzle forward (bimanual handover) 63%, Open the toolbox with a screwdriver 16%, Clamp electrical component 4%.Sort screws into binse01Datum-corner alignmente02Pour screws into the funnele03Connect DC jacke04Route cable through closed hoopse05Put tools in standing toolboxe06Place loaded boxes in container (bimanual)e07Place spray bottles upright, nozzle forward (bimanual handover)e08Open the toolbox with a screwdrivere09Clamp electrical componente10Basic skillsHigh precisionBimanualTool use

Training parameters

Starting checkpoint
From scratch
Frozen part
None
Total batch size
256
Learning rate
2e-4
Schedule
Warmup, then cosine to 0
Action chunk
32
Action space
Absolute TCP pose

Compare every model's parameters

Rollouts

Mid-training dataM0
Task-specific dataD300
SpatialNominal

One clip per environment from this entry's evaluation, beside where each evaluation episode placed the environment's objects, coloured by its rating; a placement several episodes shared is drawn once, in their ratings' colours mixed. Use the button beside each environment to open its evaluation dataset in the dataset viewer. One episode is available to watch; the remaining episodes can be requested in case of a dispute or for another legitimate reason.

  1. Sort screws into bins
    Success
    0%
    Progress
    44%
  2. Pour screws into the funnel
    Success
    87%
    Progress
    90%
  3. Connect DC jack
    Success
    0%
    Progress
    13%
  4. Route cable through closed hoops
    Success
    0%
    Progress
    39%
  5. Open the toolbox with a screwdriver
    Success
    30%
    Progress
    51%
  6. Held-out benchmark environment
    Datum-corner alignment
    Success
    27%
    Progress
    27%
  7. Held-out benchmark environment
    Put tools in standing toolbox
    Success
    3%
    Progress
    22%
  8. Held-out benchmark environment
    Place loaded boxes in container (bimanual)
    Success
    73%
    Progress
    83%
  9. Held-out benchmark environment
    Place spray bottles upright, nozzle forward (bimanual handover)
    Success
    100%
    Progress
    100%
  10. Held-out benchmark environment
    Clamp electrical component
    Success
    3%
    Progress
    42%

Success by dataset size

e01 · Sort screws into bins
D10: 3% (95% CI 1%–11%)D100: 0% (95% CI 0%–6%)D300: 2% (95% CI 0%–9%)D10D100D300
e02 · Datum-corner alignment
D10: 2% (95% CI 0%–9%)D100: 27% (95% CI 17%–39%)D300: 22% (95% CI 13%–34%)D10D100D300
e03 · Pour screws into the funnel
D10: 12% (95% CI 6%–22%)D100: 57% (95% CI 44%–68%)D300: 77% (95% CI 65%–86%)D10D100D300
e04 · Connect DC jack
D10: 0% (95% CI 0%–6%)D100: 5% (95% CI 2%–14%)D300: 0% (95% CI 0%–6%)D10D100D300
e05 · Route cable through closed hoops
D10: 0% (95% CI 0%–6%)D100: 0% (95% CI 0%–6%)D300: 0% (95% CI 0%–6%)D10D100D300
e06 · Put tools in standing toolbox
D10: 0% (95% CI 0%–6%)D100: 2% (95% CI 0%–9%)D300: 2% (95% CI 0%–9%)D10D100D300
e07 · Place loaded boxes in container (bimanual)
D10: 2% (95% CI 0%–9%)D100: 32% (95% CI 21%–44%)D300: 63% (95% CI 51%–74%)D10D100D300
e08 · Place spray bottles upright, nozzle forward (bimanual handover)
D10: 3% (95% CI 1%–11%)D100: 90% (95% CI 80%–95%)D300: 95% (95% CI 86%–98%)D10D100D300
e09 · Open the toolbox with a screwdriver
D10: 2% (95% CI 0%–9%)D100: 18% (95% CI 11%–30%)D300: 28% (95% CI 19%–41%)D10D100D300
e10 · Clamp electrical component
D10: 0% (95% CI 0%–6%)D100: 12% (95% CI 6%–22%)D300: 2% (95% CI 0%–9%)D10D100D300

Success by spatial configuration

e01 · Sort screws into bins
Nominal: 2% (95% CI 1%–8%)Interpolation: 1% (95% CI 0%–6%)ComingsoonNominalInterp.Extrap.
e02 · Datum-corner alignment
Nominal: 21% (95% CI 14%–31%)Interpolation: 12% (95% CI 7%–21%)ComingsoonNominalInterp.Extrap.
e03 · Pour screws into the funnel
Nominal: 57% (95% CI 46%–66%)Interpolation: 40% (95% CI 30%–50%)ComingsoonNominalInterp.Extrap.
e04 · Connect DC jack
Nominal: 2% (95% CI 1%–8%)Interpolation: 1% (95% CI 0%–6%)ComingsoonNominalInterp.Extrap.
e05 · Route cable through closed hoops
Nominal: 0% (95% CI 0%–4%)Interpolation: 0% (95% CI 0%–4%)ComingsoonNominalInterp.Extrap.
e06 · Put tools in standing toolbox
Nominal: 2% (95% CI 1%–8%)Interpolation: 0% (95% CI 0%–4%)ComingsoonNominalInterp.Extrap.
e07 · Place loaded boxes in container (bimanual)
Nominal: 36% (95% CI 26%–46%)Interpolation: 29% (95% CI 21%–39%)ComingsoonNominalInterp.Extrap.
e08 · Place spray bottles upright, nozzle forward (bimanual handover)
Nominal: 66% (95% CI 55%–75%)Interpolation: 60% (95% CI 50%–70%)ComingsoonNominalInterp.Extrap.
e09 · Open the toolbox with a screwdriver
Nominal: 20% (95% CI 13%–29%)Interpolation: 12% (95% CI 7%–21%)ComingsoonNominalInterp.Extrap.
e10 · Clamp electrical component
Nominal: 9% (95% CI 5%–17%)Interpolation: 0% (95% CI 0%–4%)ComingsoonNominalInterp.Extrap.