20 trials
Arrow-key sequence
Press LEFT, DOWN, then RIGHT with three assigned fingers.
keypress-ldrSetup guide ↗RC DexBench · Community results
Dexterous hands, measured in the same tasks.
Run a fixed protocol, publish the result, and compare hardware and policies across the same benchmark.
Community leaderboard
Compare results
Public submissions are community reported. Demo rows use illustrative scores to show the ranking format; they are not measured results.
| Rank | Team | Hardware | Policy | Success rate | Key accuracy | Valid trials | Mean time | Evidence |
|---|---|---|---|---|---|---|---|---|
| Loading public results… | ||||||||
Demo scores are illustrative only and are not measured DexBench results.
Rankings are calculated within a task and protocol version. The task-specific score reports key sequence accuracy, piano note accuracy and the share of notes played at MIDI velocity 50–90, or pick-and-place success by object. Mean completion time breaks success-rate ties. Custom tasks are ranked only against the same task ID and protocol version. Results have not been independently verified unless specifically labeled.
Three fixed tasks
One protocol per task
20 trials
Piano sequence
Play MIDI notes 60, 62, then 64 with even, medium force.
piano-seqSetup guide ↗
30 trials · 10 per object
Pick and place
Move a ball, cube, or marker from zone A into the tray.
pick-place-abSetup guide ↗Contribute a run
Put your result on the board.
Upload the results.json written by DexBench, describe the hardware and policy, and link the public trial recordings in DexData. The leaderboard refreshes after an accepted submission.
Standard runs use protocol version 0.1.0 and the full trial count shown above. Custom tasks must include a public task specification link and a complete fixed trial count.
Top-ranked contributors may receive a discount on a future Robotics Center purchase. We’ll contact eligible contributors who opt in; no discount amount is promised here.
Full submission format ↗