ALMI-X Whole-Body Motion Dataset
Simulated corpus combining robotic data (complete trajectories) and linguistic descriptions generated for each combined movement.
Several thousand trajectories in .npy and texts in .tar.gz (about 200 steps per episode)
Apache 2.0
Description
ALMI-X is a large-scale multimodal dataset simulated in MuJoCo, designed to train humanoid motion control models. It combines lower-body locomotion commands and complex upper-body movements, with each combination associated with an automatically generated text description. The data includes robot states, actions, joint positions, global translations, and rotations, over approximately 200 steps per trajectory (4 seconds).
What is this dataset for?
- Training robotic agents in complex locomotion and the imitation of movements
- Make the link between textual commands and physical actions (language-movement learning)
- Test simulated reinforcement learning policies before actual deployment
Can it be enriched or improved?
Yes, the dataset can be enriched by combining other movement styles, adapting text commands to other languages, or by adding additional sensors (visual or haptic). It is also possible to adapt it to other simulation environments or to physical robots.
🔎 In summary
🧠 Recommended for
- Robotics researchers
- Imitator movement projects
- Language-control modeling
🔧 Compatible tools
- MuJoco
- NumPy
- PyBullet
- PyTorch RL
- Stable Baselines
💡 Tip
Use the various speed/direction levels to create progressively difficult tasks in your RL experiences.
Frequently Asked Questions
Can this dataset be used to train a real robot?
Yes, the trajectories are designed to be transferable from the MuJoCo simulator to real robotic platforms.
Are text descriptions natural or automatically generated?
They are automatically generated according to a precise pattern linking type of movement, direction and speed level.
Do we need simulation knowledge to use this dataset?
Yes, a basic understanding of simulation in MuJoCo and data processing in .npy is recommended.
