README.md

metadata

pretty_name: TeleArms raw demonstrations (three tasks)
tags:
  - robotics
extra_gated_prompt: Access is granted manually to collaborators of the TeleArms project.

TeleArms raw demonstrations: three tasks

Raw human teleoperation demonstrations of a simulated bimanual YAM robot (MuJoCo), recorded in the TeleArms game. Every demo here succeeded and passed the acceptance checks of the TeleArms v5 training dataset; this repo holds all of that dataset's human demos of three tasks:

• Red Cube In Front Of Blue Cube: 6,109 demos
• L Cube Arrangement: 11,990 demos
• Three Cube Pyramid Stacking: 2,547 demos

20,646 demos, 246 GB, in 50 tar shards.

Files

data/<task>/<task>-NNN.tar holds <task>/<uuid>/ per demo:

• <uuid>.safetensors (played in the cloud) or <uuid>.safetensors.gz (played in the browser, gzip-compressed): the recorded robot and object states (info_qpos, info_qvel) and the commanded actions (info_demo_action) at 500 Hz, with metadata (environment name, seed, package versions; for browser play also the start layout reset_layout).
• <uuid>.json: the game's record of the attempt (task, duration, success, metrics).
• verification.json (browser play only): the server's replay check of the demo.

manifest.csv lists every demo with its task, where it was played (cloud or local = browser) and its shard.

Loading

The trajectories are read and replayed with the roboeval package of the TeleArms code base: Demo.from_safetensors(path) reads both file types (a gzipped file must keep its .gz suffix). A browser-play demo replays only after its reset_layout is applied on top of the seeded reset; the seed alone does not reproduce its start layout.