Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

DROID

Benchmarks

Task NameDataset NameSOTA ResultTrend
Multi-step manipulationDROID Tabletop Multi-step tasks
Success Rate98
18
Semantic reasoning manipulationDROID Tabletop Semantic tasks
Success Rate26
18
Rearrangement with distractorsDROID Tabletop Distractor tasks
Success Rate27
18
Video GenerationDROID Wrist
LPIPS0.364
16
Offline Action MatchingDROID
Action Score62.1
14
Predicate DetectionDROID phasic episodes (test)
F1 Score96.8
12
Pick-and-placeDROID Tabletop Simple tasks
Success Rate27
12
ClassificationDroid (test)
Accuracy97.18
10
Physical RoutingDROID memory axis (physical evaluation)
Q (%)100
10
Video GenerationDROID (Exterior)
LPIPS0.117
8
Future state retrievalDROID t+5 (2.0s prediction horizon)
Hit@142.97
8
Future state retrievalDROID t+3 (1.2s prediction horizon)
Hit@148.27
8
Future state retrievalDROID t+1 (0.4s prediction horizon)
Hit@154.11
8
Video DepthDROID
Abs Rel0.223
8
Policy IdentificationDROID Specialization
Probe Accuracy100
7
Robot Policy LearningDROID Franka Panda
Average Success Rate47.4
7
Video-to-Video GenerationDroid (test)
VBench0.81
6
Interactive long-trajectory generationDROID (val)
PSNR23.56
6
Video Frame Rank-CorrelationDROID
VOC Rank-Correlation (Sparse)0.99
6
Video GenerationDROID
PSNR24.37
5
Robotic ManipulationDROID Sharpa Hand Unseen v1 (test)
Pick and Place Progression Rate93.8
5
Robotic ManipulationDROID YAM Arm and Gripper Unseen v1 (test)
Pick and Place Progression Rate93.8
5
Robotic ManipulationDROID UMI Gripper Unseen v1 (test)
Pick and Place Progression Rate97.9
5
Robotic ManipulationDROID Original Gripper v1 (test)
Pick and Place Progression Rate100
5
Wipe Table with TowelReal-world DROID platform
Success Rate95
5
Showing 25 of 53 rows