szc97/FLORES-HIMLoco-Policy

FLORES HIMLoco Policy

This repository provides a pre-trained HIM-based locomotion policy checkpoint for FLORES, a reconfigured wheel-legged quadrupedal robot for enhanced steering and adaptability.

Paper: A Reconfigured Wheel-Legged Robot for Enhanced Steering and Adaptability
GitHub: https://github.com/ZhichengSong6/FLORES

Checkpoint

The provided checkpoint is:

May22_0943.pt

This policy was trained for the FLORES mdog robot using the HIMLoco-based reinforcement learning pipeline.

Observation and Action Space

The policy uses a 12-frame history of one-step observations. Each one-step observation has 57 dimensions:

[0:3]    base angular velocity
[3:6]    projected gravity
[6:9]    velocity command: linear x, linear y, yaw
[9:25]   joint position error
[25:41]  joint velocity
[41:57]  previous action

The full policy observation has:

57 × 12 = 684 dimensions

The latest one-step observation is placed at the front of the history buffer.

During training, the privileged observation has 250 dimensions:

[0:57]     one-step policy observation
[57:60]    base linear velocity
[60:63]    external disturbance force
[63:250]   height measurements

The policy outputs 16 actions. The action order follows the FLORES / mdog joint order:

[0]   FL_hip_joint
[1]   FL_thigh_joint
[2]   FL_calf_joint
[3]   FL_ankle_joint
[4]   FR_hip_joint
[5]   FR_thigh_joint
[6]   FR_calf_joint
[7]   FR_ankle_joint
[8]   RL_hip_joint
[9]   RL_thigh_joint
[10]  RL_calf_joint
[11]  RL_ankle_joint
[12]  RR_hip_joint
[13]  RR_thigh_joint
[14]  RR_calf_joint
[15]  RR_ankle_joint

The action scale is:

action_scale = 0.25

Please verify the actual DOF order from the loaded URDF / Isaac Gym asset before deploying the policy on a modified robot model.

Training Code

The FLORES environment files for HIMLoco are available in the GitHub repository:

code/HIM_FLORES/

These files include:

mdog_config.py
mdog_robot.py

To use them, first install HIMLoco, then place the two files under:

HIMLoco/legged_gym/legged_gym/envs/mdog/

Please refer to the GitHub README for detailed setup, task registration, and training instructions.

Robot Asset

The policy is trained for the FLORES mdog model. The expected robot asset path in the training configuration is:

HIMLoco/legged_gym/resources/robots/mdog/urdf/mdog.urdf

If the robot asset is placed elsewhere, please modify the asset path in mdog_config.py.

Notes

• This .pt file is a pre-trained policy checkpoint used with the HIMLoco / legged_gym training setup.
• The current environment uses 16 actions corresponding to the robot joints.
• The checkpoint is intended to be used together with the FLORES GitHub repository and HIMLoco.
• If the robot model, observation structure, or terrain sensing setup is modified, the policy may need to be retrained or adjusted.

Citation

If you find this work useful, please cite:

@ARTICLE{flores2026,
  author={Song, Zhicheng and Xu, Jinglan and Zheng, Chunxin and Li, Yulin and Bi, Zhihai and Ma, Jun},
  journal={IEEE Robotics and Automation Letters}, 
  title={A Reconfigured Wheel-Legged Robot for Enhanced Steering and Adaptability}, 
  year={2026},
  volume={11},
  number={6},
  pages={7444-7451},
  keywords={Legged robots;field robots},
  doi={10.1109/LRA.2026.3688384}}