---
license: mit
pipeline_tag: robotics
---

# TokenHSI Checkpoint Instruction

Please download the checkpoints and organize them as the following file structure:

```
|-- tokenhsi
|-- lpanlib
|-- body_models
    |-- smpl
        |-- SMPL_FEMALE.npz
        |-- SMPL_MALE.npz
        |-- SMPL_NEUTRAL.pkl
        |-- ...
|-- output
    |-- single_task
        |-- ckpt_traj.pth
        |-- ckpt_sit.pth
        |-- ckpt_climb.pth
        |-- ckpt_carry.pth 
    |-- tokenhsi
        |-- ckpt_stage1.pth
        |-- ...
        |-- (more models are coming soon!)
```

* We first provide four single-task environments, including path-following, sitting, climbing, and carrying. For each task, we train a specialized policy using AMP. The trained models are:
  *   single_task/ckpt_traj.pth
  *   single_task/ckpt_sit.pth
  *   single_task/ckpt_climb.pth
  *   single_task/ckpt_carry.pth

* We then provide our multi-task environment for training the versatile, unified transformer policy. The trained model is:
  *  tokenhsi/ckpt_stage1.pth

* Finally, we perform policy adaptation to generalize the learned skills to novel tasks. We list the trained models across various downstream tasks.
  * Skill Composition
     * Follow + Carry: tokenhsi/ckpt_stage2_comp_traj_carry.pth
     * Sit + Carry: tokenhsi/ckpt_stage2_comp_sit_carry.pth
     * Climb + Carry: tokenhsi/ckpt_stage2_comp_climb_carry.pth
  * Object Shape Variation
     * Chair: tokenhsi/ckpt_stage2_objShape_chair.pth
     * Table: tokenhsi/ckpt_stage2_objShape_table.pth
  * Terrain Shape Variation
     * Follow: tokenhsi/ckpt_stage2_terrainShape_traj.pth
     * Carry: tokenhsi/ckpt_stage2_terrainShape_carry.pth
  * Long-horizon Task Completion 
     * tokenhsi/ckpt_stage2_long_task.pth

github repo: https://github.com/liangpan99/TokenHSI  
arXiv paper: https://arxiv.org/abs/2503.19901