---
license: mit
task_categories:
  - robotics
tags:
  - robotics
  - lerobot
  - robosuite
  - mimicgen
  - smolvla
---

# Robosuite MimicGen 100 — State-Matched Recorded Dataset

This dataset contains 300 episodes (100 per task) of scripted expert demonstrations for three robosuite manipulation tasks, recorded by **replaying mimicgen HDF5 demos in a matched simulation environment**.

## Tasks

| Task | Episodes | Frames | Episode Length | Success Rate |
|------|----------|--------|---------------|-------------|
| Coffee | 100 | 22,225 | 210–249 (mean 222) | 100/100 |
| Square_D0 | 100 | 15,328 | 135–172 (mean 153) | 100/100 |
| Stack | 99 | 10,912 | 82–131 (mean 110) | 96/100 |
| **Total** | **299** | **48,465** | | **296/300** |

*Note: 1 Stack episode has 0 frames (recording edge case), and 4 Stack demos failed in the original HDF5.*

## Features

| Key | Shape | Type | Description |
|-----|-------|------|-------------|
| `observation.images.image` | (3, 256, 256) | video | Agentview camera |
| `observation.images.wrist_image` | (3, 256, 256) | video | Robot wrist (eye-in-hand) camera |
| `observation.state` | (8,) | float32 | `[x, y, z, axisangle_x, axisangle_y, axisangle_z, gripper_q1, gripper_q2]` |
| `action` | (7,) | float32 | `[x, y, z, axisangle_x, axisangle_y, axisangle_z, gripper]` (OSC_POSE controller, range [-1, 1]) |

- **FPS:** 20
- **Robot:** Panda
- **Controller:** OSC_POSE (Operational Space Control, 6-DOF pose + gripper)
- **Camera resolution:** 256×256
- **Format:** LeRobot v3.0 (video-encoded with SVT-AV1)

## How This Dataset Was Created

### Source Data

The original demonstrations come from the **mimicgen** project's core HDF5 datasets, hosted on HuggingFace:
- **Repository:** [`amandlek/mimicgen_datasets`](https://huggingface.co/datasets/amandlek/mimicgen_datasets)
- **Files used:** `core/coffee_d0.hdf5`, `core/square_d0.hdf5`, `core/stack_d0.hdf5`
- Each HDF5 contains 100 scripted expert episodes with `states`, `actions`, `obs`, `rewards`, and `dones`.

### Why Not Use Existing LeRobot Datasets?

Existing lerobot-format robosuite datasets (e.g., `locht131/mimicgen_100_lerobot`) were created by **replaying a trained BC policy** in the environment — not by copying the raw HDF5 scripted actions. This means:

1. **Actions differ significantly** from the ground-truth scripted demos (different magnitudes, sometimes negative correlations)
2. The BC policy may have used a **different env configuration** (controller, camera, robosuite version)
3. Training on these datasets and evaluating in a different env configuration leads to **0% success rate** even after overfitting

### Conversion Process

We recorded a fresh dataset by **replaying the exact HDF5 scripted actions in our evaluation environment**, using matching initial states:

1. **Download** the mimicgen core HDF5 files from `amandlek/mimicgen_datasets`
2. For each demo episode:
   - Create a robosuite env with our exact eval configuration:
     ```python
     robosuite.make(
         env_name=task,           # e.g. "Coffee"
         robots="Panda",
         has_offscreen_renderer=True,
         use_camera_obs=True,
         use_object_obs=False,
         camera_names=["agentview", "robot0_eye_in_hand"],
         camera_heights=256,
         camera_widths=256,
         control_freq=20,
         controller_configs=robosuite.load_controller_config(default_controller="OSC_POSE"),
     )
     ```
   - Reset the env, then **set the Mujoco simulation state** to match the HDF5 demo's initial state:
     ```python
     sim = env.sim
     state_obj = sim.get_state()
     state_obj.qpos[:] = hdf5_states[0, :sim.model.nq]
     state_obj.qvel[:] = hdf5_states[0, sim.model.nq:sim.model.nq+sim.model.nv]
     sim.set_state(state_obj)
     sim.forward()
     ```
   - Step through the demo's actions, recording `(observation.state, action, images)` at each timestep
3. Save as lerobot v3.0 dataset using `LeRobotDataset.create()` with video encoding

**Key insight:** Setting `sim.set_state()` from the HDF5 ensures the scripted actions produce the intended trajectories. Without this, env randomization means the actions are essentially random from the env's perspective.

### Conversion Script

The recorder is at `scripts/record_from_hdf5.py` in the [active_learning_vlas](https://github.com/utiasDSL/active_learning_vlas) repo (`robosuite` branch).

To reproduce:
```bash
python scripts/record_from_hdf5.py \
    --tasks Coffee Square_D0 Stack \
    --episodes_per_task 100 \
    --output_dir data/mimicgen_100_recorded
```

Requires: `robosuite`, `mimicgen`, `lerobot`, GPU with EGL support for rendering.

## Usage with LeRobot

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("dlsmarta/robosuite_mimicgen_100")
print(f"Episodes: {dataset.num_episodes}, Frames: {dataset.num_frames}")
```

Or in a training config:
```yaml
dataset:
  repo_id: dlsmarta/robosuite_mimicgen_100
  video_backend: pyav
```

## Citation

If you use this dataset, please also cite the original mimicgen datasets:

```bibtex
@inproceedings{mimicgen2023,
    title={MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations},
    author={Mandlekar, Ajay and Nasiriany, Soroush and Wen, Bowen and Akinola, Iide and Kannan, Animesh and Sheridan, Jackson and Azara, Andrew and Stone, Travis and Fan, Tse and Shridhar, Mohit and others},
    booktitle={7th Annual Conference on Robot Learning},
    year={2023}
}
```