---
license: apache-2.0
task_categories:
- reinforcement-learning
- robotics
tags:
- Minari
- TeleArms
- bimanual
- simulation
- offline-rl
---
# TeleArms offline RL datasets (state)

30 [Minari](https://minari.farama.org) datasets of human teleoperated demonstrations of 10 bimanual
YAM cube tasks in MuJoCo: each task in three subsets of 1,500 demos, picked at random
(`random-1500`), by player ELO (`elo-1500`) and by demo score (`score-1500`). The ids are
`telearms/<Task>/<subset>-v0`:

```python
import os
import minari
from huggingface_hub import snapshot_download

root = os.environ.get("MINARI_DATASETS_PATH", os.path.expanduser("~/.minari/datasets"))
snapshot_download("mliu48/telearms", repo_type="dataset", local_dir=os.path.join(root, "telearms"),
                  allow_patterns="CubePatchPlacement/random-1500-v0/*")
dataset = minari.load_dataset("telearms/CubePatchPlacement/random-1500-v0")
env = dataset.recover_environment()  # needs telearms-gym installed
```

The environment is `TeleArmsEnv(task, obs_mode="privileged")` in
[telearms-gym](https://github.com/liuwilli10101/telearms-gym) (`gymnasium_env/telearms_env.py`). On a machine without EGL, set `MUJOCO_GL=disable`
before importing it.

**Observations** are the simulator's state, not images: the robot's 14 joint and gripper values,
then each cube's position, orientation and velocity (13 numbers per cube), then each goal patch's
position and orientation (7 per patch).

| Task | Observation size |
|---|---|
| CubePatchPlacement | 54 |
| FrontEdgeCubeRow | 53 |
| FrontToBackCubeLine | 53 |
| LCubeArrangement | 66 |
| LetterFaceAlignment | 27 |
| RedCubeInFrontOfBlueCube | 40 |
| RedLeftOfBluePlacement | 40 |
| SemanticCubeOnRedPatchAUp | 34 |
| SequentialPushBlockToTarget | 54 |
| ThreeCubeWallRow | 53 |

**Rewards.** Every demo was replayed in its task's environment and scored by its rules. `rewards`
is the sparse reward, 1.0 on the step that reaches success, else 0. `infos["reward_dense"]` is the
environment's `reward_type="dense"` reward: 1/N on the first step that reaches each of the task's
N stages (grasp, lift, place, ...) plus 1.0 on success, at most 2 per episode. Each demo ends at
its first success (`terminations`), as the environment ends an episode. Of the 45,000 demos,
665 (1.5%) were accepted when played but never satisfy the environment's task check; they end with
`truncations` and no reward.

The same demos with camera images are the `offline-rl` branch of
`mliu48/telearms-yam-cube-v5-exp-<subset>`:
`LeRobotDataset("mliu48/telearms-yam-cube-v5-exp-random-1500", revision="offline-rl")`.