<!-- BEGIN auto-index -->
<!-- Regenerated by scripts/coffee/upload_dataset_to_hub.py — edits here are overwritten. -->

**Domain:** simulation · **Environments:** **robosuite, single 6-DoF SO-101 arm** (`env-sim-robosuite-so101`), **Simulated 17-DoF XLeRobot platform** (`env-sim-xlerobot`) · **Format:** LeRobot `v3.0`

Layout is **`<environment>/<task>/`** — every leaf is a complete LeRobot dataset root.

| Folder | Task | Kind | Description | Environment | Episodes | Frames | FPS | Robot |
|---|---|---|---|---|---|---|---|---|
| `sim-robosuite-so101/t1_place_cup` | t1 | manipulation | Place the cup on the machine platform | robosuite, single 6-DoF SO-101 arm (`sim-robosuite-so101`) | 50 | 10,231 | 20 | `robosuite_sim` |
| `sim-robosuite-so101/t3_cup_to_tray` | t3 | manipulation | Move the filled cup from the machine to the tray | robosuite, single 6-DoF SO-101 arm (`sim-robosuite-so101`) | 50 | 19,485 | 20 | `robosuite_sim` |
| `sim-robosuite-so101/t5_tray_to_table` | t5 | manipulation | Move the filled cup from the tray to the user's table | robosuite, single 6-DoF SO-101 arm (`sim-robosuite-so101`) | 50 | 17,567 | 20 | `robosuite_sim` |
| `sim-xlerobot/t1_place_cup` | t1 | manipulation | Place the cup on the machine platform | Simulated 17-DoF XLeRobot platform (`sim-xlerobot`) | 51 | 17,988 | 20 | `xlerobot_sim` |
| `sim-xlerobot/t2_push_button` | t2 | manipulation | Push the machine's brew button | Simulated 17-DoF XLeRobot platform (`sim-xlerobot`) | 50 | 7,050 | 20 | `xlerobot_sim` |
| `sim-xlerobot/t3_cup_to_tray` | t3 | manipulation | Move the filled cup from the machine to the tray | Simulated 17-DoF XLeRobot platform (`sim-xlerobot`) | 50 | 19,209 | 20 | `xlerobot_sim` |
| `sim-xlerobot/t5_tray_to_table` | t5 | manipulation | Move the filled cup from the tray to the user's table | Simulated 17-DoF XLeRobot platform (`sim-xlerobot`) | 50 | 17,250 | 20 | `xlerobot_sim` |

**Not yet uploaded:** `t4_navigate` (t4).

### Per-frame features

| Feature | dtype | shape | Meaning |
|---|---|---|---|
| `action` | float32 | `[6]` | Commanded targets — 17-D on XLeRobot (left/right arm ×6, head ×2, base x/y/theta vel); 6-D on SO-101 (shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper). |
| `observation.state` | float32 | `[117]` | RAW telemetry — 84-D on XLeRobot (17 motors × {pos, current_raw, load_raw, vel_hw} + base x/y/theta vel + 13 GY-91 IMU channels; imu.mag_* is NaN, an MPU6500 has no magnetometer). 117-D on SO-101/sim, which additionally bakes in the EKF breakdown (estimated/tau_ext, model/tau_*, tcp/wrench/*, hpi/gripper/*). |
| `observation.sim_model_force` | float32 | `[36]` | Ground truth, sim only (36-D): per-joint model decomposition (tau_motor, tau_gravity, tau_coriolis, tau_inertial, tau_friction, tau_model). |
| `observation.sim_contact_force` | float32 | `[24]` | Ground truth, sim only (24-D): per-joint true external torque (tau_ext, tau_ext_residual, tau_ext_jac, tau_ext_contact) — the label an estimator is graded against. |
| `observation.sim_tcp_wrench` | float32 | `[12]` | Ground truth, sim only (12-D): true TCP wrench, contact and F/T-sensor variants. |
| `observation.sim_pose` | float32 | `[7]` | Ground truth, sim only (7-D): cup position + quaternion. |
| `observation.sim_task` | float32 | `[2]` | Ground truth, sim only (2-D): task success / failure flags. |
| `observation.hpi` | float32 | `[9]` | Force channel (9-D): [grip tau_ext, grip q, grip dq, TCP wrench fx fy fz tx ty tz] — input to the force-aware (A1) models. NOT present on XLeRobot platform data: estimation there is a deliberate offline step so the estimator stays an experimental axis. |
| `observation.images.top` | video | `[480, 640, 3]` | Head / overview camera (RGB). |
| `observation.images.top_depth` | image | `[480, 640, 1]` | Overview depth (int16, 1-channel). Not used by ACT; dropped when building training variants. |
| `observation.images.head` | video | `[480, 640, 3]` | Head / overview camera (RGB). |
| `observation.images.head_depth` | image | `[480, 640, 1]` | Overview depth (int16, 1-channel). Not used by ACT; dropped when building training variants. |

Load one leaf (`allow_patterns` matters — without it you pull the whole repo):

```python
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

repo, leaf = "IntelligentDecisionLab/xlerobot-coffee-sim", "sim-robosuite-so101/t1_place_cup"
root = snapshot_download(repo, repo_type="dataset", allow_patterns=f"{leaf}/*")
ds = LeRobotDataset(repo, root=f"{root}/{leaf}")
```
<!-- END auto-index -->