DexArena demonstrations

Teleoperated demonstrations for the DexArena benchmark (github.com/SIPAIR-Elastic/DexArena): the 30 stage tasks of its 10 long-horizon dexterous manipulation tasks, recorded in simulation with the Sharpa hand on a Franka arm, as LeRobot v3 datasets. <hand> is the env id's end effector in lower case (sharpa, wuji, xhand, gripper), so one hand's datasets are one directory; the task table with every env id is in the code repository's README.

• data/<hand>/<ENV_ID>/ — the release set: 100 replay-successful episodes per stage task, with multi-view videos. See data/README.md.
• data_append/<hand>/<ENV_ID>/ — the remaining reviewed episodes of each stage task, state only. See data_append/README.md.
• assets/task_object_meshes/ — the object meshes the environments load at run time.
• ckpt/, model/ — baseline checkpoints.

Download from the code repository's root so the paths match its tools:

hf download SIPAIR/DexArena --repo-type dataset --include "assets/**" --local-dir tasks
hf download SIPAIR/DexArena --repo-type dataset --include "data/<hand>/<ENV_ID>/**" --local-dir .

This is the release the DexArena paper reports on. The same episodes were first uploaded under earlier env ids in a data/sharpa_hand/env_* layout; that layout, together with the data_no_randomization/ and data_0824/ folders, remains in the repository history at commit 178bf16ce851 and is not maintained. Every raw recording under replay/source_*/ is byte-identical between the two, and identity_manifest.recorded_as in each episode's meta.json keeps the env id, task name and code fingerprint it was recorded under.