ALMI-X

🌍website  📊code  📖paper

Overview

We release a large-scale whole-body motion control dataset - ALMI-X, featuring high-quality episodic trajectories from MuJoCo simulations deployable on real robots, based on our humanoid control policy - ALMI.

Dataset Instruction

We collect ALMI-X dataset in MuJoCo simulation by running the trained ALMI policy. In this simulation, we combine a diverse range of upper-body motions with omnidirectional lower-body commands, and employ a pre-defined paradigm to generate corresponding linguistic descriptions for each combination. (i) For the upper-body, we collect data using our upper-body policy to track various motions from a subset of the AMASS dataset, where we remove entries with indistinct movements or those that could not be matched with the lower-body commands, such as push from behind. (ii) For the lower-body, we first categorize command directions into several types according to different combination of linear and angular velocity command and define 3 difficulty levels for command magnitudes, then the lower-body command is set by combining direction types and difficulty levels. Overall, each upper-body motion from the AMASS subset is paired with a specific direction type and a difficulty level serving as the inputs of policy to control the robot. In addition, trajectories in which the lower body stand still while the upper body tracks motions are also incorporated into the dataset. Each language description in ALMI-X is organized as "[movement mode] [direction] [velocity level] and motion`"}, each of which corresponds to the data collected from a trajectory lasting about 4 seconds with 200 steps. For each trajectory$, we run two policies (i.e., lower policy and upper policy) based on the commands obtained from the aforementioned combinations to achieve humanoid whole-body control.

How to Use Dataset

• We release all of the text description data text.tar.gz; the trajectory data data.tar.gz with robot states, actions, DoF position, global position and global orientation informations.
• We release the train set split train.txt

Here we offer a simple demo code to introduce the data formats in the dataset:

import numpy as np

data = np.load("data_path"+"/xxx.npy", allow_pickle=True)
data.item()['obs'] # [frame_nums, 71]
data.item()['actions'] # [frame_nums, 21]
data.item()['dof_pos'] # [frame_nums, 21]
data.item()['root_trans'] # [frame_nums, 3]
data.item()['root_rot'] # [frame_nums, 4]

Dataset Statistics

Percentage of steps for different categories of motions before and after data augmentation.
The visualization of $x-y$ coordinates of the robot for each step in the dataset. We down-sample the data for visualization.

Dataset Collection Pipeline

We release our datase collection code at our github repository: Data_Collection.

Citation

If you find our work helpful, please cite us:

      @misc{shi2025almi,
        title={Adversarial Locomotion and Motion Imitation for Humanoid Policy Learning}, 
        author={Jiyuan Shi and Xinzhe Liu and Dewei Wang and Ouyang Lu and Sören Schwertfeger and Fuchun Sun and Chenjia Bai and Xuelong Li},
        year={2025},
        eprint={2504.14305},
        archivePrefix={arXiv},
        primaryClass={cs.RO},
        url={https://arxiv.org/abs/2504.14305}, 
  }
}