NOVA Dataset - Reachy 2 Pick-and-Place Demonstrations

Expert demonstration dataset for training vision-language-action models on Pollen Robotics' Reachy 2 humanoid robot. Collected in MuJoCo simulation with domain randomization.

Dataset Description

This dataset contains 100 episodes of pick-and-place manipulation tasks, designed for fine-tuning NVIDIA GR00T N1.6 or other imitation learning policies.

Key Features

Feature	Description
Format	LeRobot v2.1 (parquet + H264 video)
Episodes	100
Task Variations	32 (4 objects × 8 colors)
Camera Views	2 (front_cam, workspace_cam)
Resolution	640×480 @ 15 FPS
Domain Randomization	Position, lighting, camera jitter

Task Description

The robot performs pick-and-place tasks with natural language instructions:

• "Pick up the red cube and place it in the box"
• "Pick up the blue cylinder and place it in the box"
• "Pick up the green capsule and place it in the box"

Dataset Structure

NOVA/
├── meta/
│   ├── info.json           # Dataset metadata
│   ├── stats.json          # Normalization statistics
│   ├── tasks.jsonl         # Task descriptions
│   └── episodes.jsonl      # Episode information
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ...
└── videos/
    └── chunk-000/
        ├── observation.images.front_cam/
        │   ├── episode_000000.mp4
        │   └── ...
        └── observation.images.workspace_cam/
            ├── episode_000000.mp4
            └── ...

Data Fields

State (Proprioception)

Field	Dimension	Description
observation.state	7	Joint positions (arm)

Joint Names:

1. shoulder_pitch (-180° to 90°)
2. shoulder_roll (-180° to 10°)
3. elbow_yaw (-90° to 90°)
4. elbow_pitch (-125° to 0°)
5. wrist_roll (-100° to 100°)
6. wrist_pitch (-45° to 45°)
7. wrist_yaw (-30° to 30°)

Action

Field	Dimension	Description
action	8	Joint positions (7 arm + 1 gripper)

Gripper: 0 = closed, 1 = open

Video

Camera	Resolution	FOV	Format
front_cam	640×480	108°	H264 MP4
workspace_cam	640×480	70°	H264 MP4

Language

Field	Description
annotation.human.task_description	Natural language task instruction

Objects and Colors

Objects (4 types)

• Cube (4cm)
• Rectangular box
• Cylinder
• Capsule

Colors (8 variations)

• Red, Green, Blue, Yellow
• Cyan, Magenta, Orange, Purple

Domain Randomization

Parameter	Range
Object position	Workspace-aware random
Lighting intensity	0.5 - 1.0
Camera jitter	±2°
Object type	Random selection
Object color	Random selection

Usage

Loading with LeRobot

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("ganatrask/NOVA")

# Access an episode
episode = dataset[0]
print(episode.keys())
# ['observation.state', 'observation.images.front_cam', 'action', ...]

Loading with HuggingFace Datasets

from datasets import load_dataset

dataset = load_dataset("ganatrask/NOVA")

Training GR00T

python -m gr00t.train \
    --dataset_repo_id ganatrask/NOVA \
    --embodiment_tag reachy2 \
    --video_backend decord \
    --num_gpus 2 \
    --batch_size 64 \
    --max_steps 30000

Collection Details

Environment

• Simulator: MuJoCo via reachy2_mujoco
• Robot: Reachy 2 humanoid (14-DOF arms, using right arm only)
• Control Frequency: 15 Hz

Collection Process

1. Random object and color selection
2. Random placement within workspace
3. Scripted expert policy execution
4. Recording of observations, states, and actions
5. Automatic episode segmentation

Collection Command

python scripts/data_collector.py \
    --episodes 100 \
    --output reachy2_dataset \
    --arm right \
    --randomize-object \
    --randomize-color \
    --cameras front_cam workspace_cam

Statistics

Statistic	Value
Total episodes	100
Avg. episode length	~150 steps
Collection rate	~2 episodes/min
Total size	~2 GB

Limitations

• Simulation only: Data collected in MuJoCo, not real robot
• Single arm: Right arm manipulation only
• Fixed task type: Pick-and-place only
• Limited objects: 4 primitive shapes

License

This dataset is released under CC-BY-4.0.

Citation

@misc{nova_dataset_2025,
  title={NOVA Dataset: Reachy 2 Pick-and-Place Demonstrations},
  author={ganatrask},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/ganatrask/NOVA}
}

Acknowledgments

• Pollen Robotics - Reachy 2 robot and MuJoCo simulation
• HuggingFace - LeRobot framework and dataset hosting
• DeepMind - MuJoCo physics engine

Related Resources

• Model: ganatrask/NOVA - Fine-tuned GR00T model
• Code: ganatrask/NOVA - Training and inference code
• Base Model: nvidia/GR00T-N1.6-3B
• LeRobot: huggingface/lerobot