DexArenaDexterous intelligence
Evaluation protocol 50 episodes per task; cross-scene scores span three held-out scenes.Read the protocol →
DexArena benchmark

Leaderboard

Skill execution · Cross-scene generalization · Long-horizon composition

Manuscript baseline results · six methods
Results
  • Six paper baselines now include all stage, cross-scene and long-horizon scores.
  • Cross-scene ± denotes standard deviation across scenes, not confidence intervals.
  • These are manuscript baselines; community submission intake is not yet open.
30 stages · training scene

Stage benchmark

Success rate over 50 episodes per stage. Each policy starts from a valid initial state for that stage.

Download CSV ↓
Average success (%)
0255075100
1GR00TVLA61.4%
2ACTImitation learning58.0%
3RLPDDemo-guided RL54.3%
4π0.5VLA52.1%
5Diffusion PolicyImitation learning39.2%
6PPODemo-guided RL27.9%

All tracks use the same 0–100% scale.

Compare each method across all three tracks

Success (%), on a shared 0–100 scale.

PPO

Stage
27.9
Scene
21.7
Sequence
0.0

RLPD

Stage
54.3
Scene
17.8
Sequence
0.0

ACT

Stage
58.0
Scene
45.4
Sequence
1.2

Diffusion Policy

Stage
39.2
Scene
17.2
Sequence
0.0

GR00T

Stage
61.4
Scene
41.7
Sequence
0.0

π0.5

Stage
52.1
Scene
49.2
Sequence
1.6

Stage performance map

0%100%

All six methods. Cells show success rate; hover for the exact value.

StagePPORLPDACTDPGR00Tπ0.5
Open Lunch Box74.0100.0100.096.0100.0100.0
Place Bread12.086.088.020.060.070.0
Scoop Tomato0.00.044.00.012.032.0
Open Cabinet62.0100.094.088.092.098.0
Pick Bottle0.014.068.018.078.054.0
Unscrew Cap0.02.062.050.062.046.0
Pour Beads0.096.0100.032.078.052.0
Open Drawer98.0100.098.098.098.098.0
Pick Charger0.062.092.00.078.078.0
Insert Plug0.00.016.012.06.012.0
Pick Bolt0.028.08.022.044.012.0
Seat Nut0.034.012.06.026.02.0
Turn Nut76.0100.096.0100.0100.0100.0
Open Laptop100.0100.0100.088.0100.082.0
Type Keys0.00.00.00.032.00.0
Click Mouse96.0100.014.092.098.062.0
Lay Flat Panel0.096.078.024.064.00.0
Insert Rail0.00.022.02.016.08.0
Fasten Screw100.0100.00.00.084.094.0
Aspirate Liquid0.00.040.02.08.026.0
Dispense Liquid0.00.00.02.08.00.0
Open Fridge90.098.0100.088.094.096.0
Apple To Blender0.00.092.042.054.038.0
Close Fridge100.0100.0100.0100.098.0100.0
Place Lid0.078.026.026.042.016.0
Press Button26.0100.066.072.064.068.0
Open Package2.092.044.02.064.058.0
Extract Apple0.042.088.022.0100.082.0
Pick Debris0.00.00.02.00.00.0
Dump Dustpan0.00.092.070.082.078.0
RankMethodFamilyAverage success (%) ↑Source
1GR00TVLA61.4Paper baseline ↗
2ACTImitation learning58.0Paper baseline ↗
3RLPDDemo-guided RL54.3Paper baseline ↗
4π0.5VLA52.1Paper baseline ↗
5Diffusion PolicyImitation learning39.2Paper baseline ↗
6PPODemo-guided RL27.9Paper baseline ↗

All 30 stages

Success rates (%). Highlighted cells are the highest result in each row.

Stage benchmark: success rates in percent
StagePPORLPDACTDiffusion PolicyGR00Tπ0.5
OpenLunchBox74.0100.0100.096.0100.0100.0
PlaceBread12.086.088.020.060.070.0
ScoopTomato0.00.044.00.012.032.0
OpenCabinet62.0100.094.088.092.098.0
PickBottle0.014.068.018.078.054.0
UnscrewCap0.02.062.050.062.046.0
PourBeads0.096.0100.032.078.052.0
OpenDrawer98.0100.098.098.098.098.0
PickCharger0.062.092.00.078.078.0
InsertPlug0.00.016.012.06.012.0
PickBolt0.028.08.022.044.012.0
SeatNut0.034.012.06.026.02.0
TurnNut76.0100.096.0100.0100.0100.0
OpenLaptop100.0100.0100.088.0100.082.0
TypeKeys0.00.00.00.032.00.0
ClickMouse96.0100.014.092.098.062.0
LayFlatPanel0.096.078.024.064.00.0
InsertRail0.00.022.02.016.08.0
FastenScrew100.0100.00.00.084.094.0
AspirateLiquid0.00.040.02.08.026.0
DispenseLiquid0.00.00.02.08.00.0
OpenFridge90.098.0100.088.094.096.0
AppleToBlender0.00.092.042.054.038.0
CloseFridge100.0100.0100.0100.098.0100.0
PlaceLid0.078.026.026.042.016.0
PressButton26.0100.066.072.064.068.0
OpenPackage2.092.044.02.064.058.0
ExtractApple0.042.088.022.0100.082.0
PickDebris0.00.00.02.00.00.0
DumpDustpan0.00.092.070.082.078.0
Average27.954.358.039.261.452.1
30 stages × 3 held-out scenes

Cross-scene

Fixed policies, no adaptation. Mean ± standard deviation across three kitchen scenes, with 50 episodes per stage per scene.

Download CSV ↓
Average success (%)
0255075100
1π0.5VLA49.2 ± 2.3%
2ACTImitation learning45.4 ± 3.6%
3GR00TVLA41.7 ± 14.2%
4PPODemo-guided RL21.7 ± 3.5%
5RLPDDemo-guided RL17.8 ± 10.9%
6Diffusion PolicyImitation learning17.2 ± 5.7%

All tracks use the same 0–100% scale. Whiskers show standard deviation across three scenes.

Compare each method across all three tracks

Success (%), on a shared 0–100 scale.

PPO

Stage
27.9
Scene
21.7
Sequence
0.0

RLPD

Stage
54.3
Scene
17.8
Sequence
0.0

ACT

Stage
58.0
Scene
45.4
Sequence
1.2

Diffusion Policy

Stage
39.2
Scene
17.2
Sequence
0.0

GR00T

Stage
61.4
Scene
41.7
Sequence
0.0

π0.5

Stage
52.1
Scene
49.2
Sequence
1.6

Stage performance map

0%100%

All six methods. Cells show scene-average success; hover for mean ± scene standard deviation.

StagePPORLPDACTDPGR00Tπ0.5
Open Lunch Box35.333.3100.035.3100.099.3
Place Bread10.00.040.08.724.071.3
Scoop Tomato0.00.023.30.06.718.7
Open Cabinet8.763.393.384.078.797.3
Pick Bottle0.00.076.04.046.069.3
Unscrew Cap0.00.04.01.39.333.3
Pour Beads0.70.042.06.016.736.0
Open Drawer97.395.382.74.090.0100.0
Pick Charger0.05.388.00.064.788.0
Insert Plug0.00.00.70.02.01.3
Pick Bolt0.70.08.72.021.318.0
Seat Nut0.00.02.70.72.03.3
Turn Nut26.711.344.011.380.087.3
Open Laptop99.348.084.788.795.366.7
Type Keys0.00.00.00.017.30.0
Click Mouse98.06.712.768.7100.064.7
Lay Flat Panel0.00.054.00.034.70.0
Insert Rail0.00.00.00.06.09.3
Fasten Screw86.070.785.30.020.094.0
Aspirate Liquid0.00.08.71.32.034.0
Dispense Liquid0.00.00.00.02.04.0
Open Fridge65.378.792.768.070.096.0
Apple To Blender0.70.071.37.318.038.7
Close Fridge100.0100.099.366.092.0100.0
Place Lid0.00.018.01.328.716.0
Press Button18.018.754.052.742.762.0
Open Package2.02.736.71.335.330.7
Extract Apple0.00.078.73.354.770.0
Pick Debris1.30.00.00.01.30.0
Dump Dustpan0.70.059.31.388.768.0
RankMethodFamilyAverage success (%) ↑Source
1π0.5VLA49.2 ± 2.3Paper baseline ↗
2ACTImitation learning45.4 ± 3.6Paper baseline ↗
3GR00TVLA41.7 ± 14.2Paper baseline ↗
4PPODemo-guided RL21.7 ± 3.5Paper baseline ↗
5RLPDDemo-guided RL17.8 ± 10.9Paper baseline ↗
6Diffusion PolicyImitation learning17.2 ± 5.7Paper baseline ↗

All 30 stages

Success rates (%). Highlighted cells are the highest result in each row.

Cross-scene: success rates in percent
StagePPORLPDACTDiffusion PolicyGR00Tπ0.5
OpenLunchBox35.3 ± 14.533.3 ± 57.7100.0 ± 0.035.3 ± 31.6100.0 ± 0.099.3 ± 1.2
PlaceBread10.0 ± 3.50.0 ± 0.040.0 ± 30.08.7 ± 4.624.0 ± 26.271.3 ± 7.0
ScoopTomato0.0 ± 0.00.0 ± 0.023.3 ± 9.00.0 ± 0.06.7 ± 6.118.7 ± 3.1
OpenCabinet8.7 ± 9.063.3 ± 54.993.3 ± 6.484.0 ± 11.178.7 ± 17.997.3 ± 2.3
PickBottle0.0 ± 0.00.0 ± 0.076.0 ± 8.04.0 ± 5.346.0 ± 36.769.3 ± 6.4
UnscrewCap0.0 ± 0.00.0 ± 0.04.0 ± 3.51.3 ± 2.39.3 ± 14.533.3 ± 11.0
PourBeads0.7 ± 1.20.0 ± 0.042.0 ± 22.36.0 ± 10.416.7 ± 17.536.0 ± 5.3
OpenDrawer97.3 ± 4.695.3 ± 4.282.7 ± 10.14.0 ± 5.390.0 ± 10.6100.0 ± 0.0
PickCharger0.0 ± 0.05.3 ± 9.288.0 ± 5.30.0 ± 0.064.7 ± 19.688.0 ± 0.0
InsertPlug0.0 ± 0.00.0 ± 0.00.7 ± 1.20.0 ± 0.02.0 ± 3.51.3 ± 1.2
PickBolt0.7 ± 1.20.0 ± 0.08.7 ± 5.02.0 ± 3.521.3 ± 13.318.0 ± 2.0
SeatNut0.0 ± 0.00.0 ± 0.02.7 ± 2.30.7 ± 1.22.0 ± 3.53.3 ± 2.3
TurnNut26.7 ± 11.011.3 ± 16.344.0 ± 19.711.3 ± 12.780.0 ± 19.787.3 ± 5.0
OpenLaptop99.3 ± 1.248.0 ± 50.184.7 ± 18.188.7 ± 3.195.3 ± 8.166.7 ± 5.0
TypeKeys0.0 ± 0.00.0 ± 0.00.0 ± 0.00.0 ± 0.017.3 ± 17.00.0 ± 0.0
ClickMouse98.0 ± 3.56.7 ± 11.512.7 ± 4.668.7 ± 15.0100.0 ± 0.064.7 ± 3.1
LayFlatPanel0.0 ± 0.00.0 ± 0.054.0 ± 27.50.0 ± 0.034.7 ± 28.40.0 ± 0.0
InsertRail0.0 ± 0.00.0 ± 0.00.0 ± 0.00.0 ± 0.06.0 ± 10.49.3 ± 3.1
FastenScrew86.0 ± 24.270.7 ± 50.885.3 ± 11.50.0 ± 0.020.0 ± 26.294.0 ± 3.5
AspirateLiquid0.0 ± 0.00.0 ± 0.08.7 ± 8.11.3 ± 1.22.0 ± 2.034.0 ± 6.9
DispenseLiquid0.0 ± 0.00.0 ± 0.00.0 ± 0.00.0 ± 0.02.0 ± 2.04.0 ± 0.0
OpenFridge65.3 ± 22.078.7 ± 35.292.7 ± 9.568.0 ± 33.370.0 ± 45.196.0 ± 3.5
AppleToBlender0.7 ± 1.20.0 ± 0.071.3 ± 14.27.3 ± 12.718.0 ± 2.038.7 ± 4.2
CloseFridge100.0 ± 0.0100.0 ± 0.099.3 ± 1.266.0 ± 47.392.0 ± 8.7100.0 ± 0.0
PlaceLid0.0 ± 0.00.0 ± 0.018.0 ± 11.11.3 ± 2.328.7 ± 12.916.0 ± 2.0
PressButton18.0 ± 5.318.7 ± 32.354.0 ± 19.352.7 ± 40.642.7 ± 10.162.0 ± 7.2
OpenPackage2.0 ± 0.02.7 ± 4.636.7 ± 22.11.3 ± 2.335.3 ± 22.130.7 ± 24.8
ExtractApple0.0 ± 0.00.0 ± 0.078.7 ± 3.13.3 ± 3.154.7 ± 39.570.0 ± 26.5
PickDebris1.3 ± 2.30.0 ± 0.00.0 ± 0.00.0 ± 0.01.3 ± 1.20.0 ± 0.0
DumpDustpan0.7 ± 1.20.0 ± 0.059.3 ± 17.91.3 ± 2.388.7 ± 2.368.0 ± 4.0
Average21.7 ± 3.517.8 ± 10.945.4 ± 3.617.2 ± 5.741.7 ± 14.249.2 ± 2.3
10 complete sequences

Long-horizon

End-to-end success over 50 episodes per task. The environment is never reset between stages. π0.5 switches stage policies; GR00T uses one model trained on pooled stage data.

Download CSV ↓
Average success (%)
0255075100
1π0.5VLA1.6%
2ACTImitation learning1.2%
3PPODemo-guided RL0.0%
4RLPDDemo-guided RL0.0%
5Diffusion PolicyImitation learning0.0%
6GR00TVLA0.0%

All tracks use the same 0–100% scale.

Compare each method across all three tracks

Success (%), on a shared 0–100 scale.

PPO

Stage
27.9
Scene
21.7
Sequence
0.0

RLPD

Stage
54.3
Scene
17.8
Sequence
0.0

ACT

Stage
58.0
Scene
45.4
Sequence
1.2

Diffusion Policy

Stage
39.2
Scene
17.2
Sequence
0.0

GR00T

Stage
61.4
Scene
41.7
Sequence
0.0

π0.5

Stage
52.1
Scene
49.2
Sequence
1.6

Sequence performance map

0%100%

All six methods. Cells show success rate; hover for the exact value.

TaskPPORLPDACTDPGR00Tπ0.5
Meal preparation0.00.00.00.00.00.0
Cabinet and pouring0.00.00.00.00.02.0
Drawer and plug0.00.00.00.00.00.0
Nut and bolt0.00.00.00.00.00.0
Office desk0.00.00.00.00.00.0
Furniture assembly0.00.08.00.00.00.0
Pipette transfer0.00.00.00.00.00.0
Juice preparation0.00.00.00.00.00.0
Package opening0.00.04.00.00.014.0
Table cleaning0.00.00.00.00.00.0
RankMethodFamilyAverage success (%) ↑Source
1π0.5VLA1.6Paper baseline ↗
2ACTImitation learning1.2Paper baseline ↗
3PPODemo-guided RL0.0Paper baseline ↗
4RLPDDemo-guided RL0.0Paper baseline ↗
5Diffusion PolicyImitation learning0.0Paper baseline ↗
6GR00TVLA0.0Paper baseline ↗

All 10 tasks

Success rates (%). Highlighted cells are the highest result in each row.

Long-horizon: success rates in percent
TaskPPORLPDACTDiffusion PolicyGR00Tπ0.5
Meal preparation0.00.00.00.00.00.0
Cabinet and pouring0.00.00.00.00.02.0
Drawer and plug0.00.00.00.00.00.0
Nut and bolt0.00.00.00.00.00.0
Office desk0.00.00.00.00.00.0
Furniture assembly0.00.08.00.00.00.0
Pipette transfer0.00.00.00.00.00.0
Juice preparation0.00.00.00.00.00.0
Package opening0.00.04.00.00.014.0
Table cleaning0.00.00.00.00.00.0
Average0.00.01.20.00.01.6

Source: DexArena manuscript, stage-level and generalization experiment tables and long-horizon appendix table. Download all results (JSON) ↓

Submit a method

Evaluation intake is not open yet.

Submission instructions will be shared when evaluation intake opens.