Hi there!
Currently I am trying to implement a well-performed locomotion policy onto a quadruped robot (unitree b2), and the robot experienced weird movement. The attempts i tried is as below:
1. URDF, the urdf is from the offical git repository, and i check the imported version, i'm sure there's no issue about the dynamics.
2. the RL policy takes things like joint position, joint velocity, target velocity of base etc as inputs and outputs the target postion of each joint, then calculates the target torque by a pd controller (i copy the params of PD controller from their gazebo demo and it performs well in isaacsim). The robot goes weird when it's in coppeliasim with bullet/mujoco/...
3. based on (2), i tried different param setups and none of those work.
Plus:
1. link of policy: https://www.dropbox.com/scl/fi/yj5nicsl ... 4ne1e&dl=0
2. link to scene https://www.dropbox.com/scl/fi/c611f8q6 ... h0dgy&dl=0
Could someone please help me with this? Thanks!
Transfer a locomotion policy of a quadruped robot from isaacsim
-
- Posts: 3
- Joined: 28 Oct 2024, 09:55
Re: Transfer a locomotion policy of a quadruped robot from isaacsim
Hello,
URDF portability across applications can be quite challenging. The URDF and your model are extremely heavy and not optimized. Also, the joint control mode should be set to force/torque in your case and many parameters adjusted, such as the armature parameter of the joints (MuJoCo). If running with other engines, you probably also have to artificially increase the inertia of most bodies.
Have a look at this Spot scene and that Spot model to get some ideas.
Cheers
URDF portability across applications can be quite challenging. The URDF and your model are extremely heavy and not optimized. Also, the joint control mode should be set to force/torque in your case and many parameters adjusted, such as the armature parameter of the joints (MuJoCo). If running with other engines, you probably also have to artificially increase the inertia of most bodies.
Have a look at this Spot scene and that Spot model to get some ideas.
Cheers
-
- Posts: 3
- Joined: 28 Oct 2024, 09:55
Re: Transfer a locomotion policy of a quadruped robot from isaacsim
These suggestions and files definately would help, thank you team!
coppelia wrote: ↑30 Oct 2024, 09:44 Hello,
URDF portability across applications can be quite challenging. The URDF and your model are extremely heavy and not optimized. Also, the joint control mode should be set to force/torque in your case and many parameters adjusted, such as the armature parameter of the joints (MuJoCo). If running with other engines, you probably also have to artificially increase the inertia of most bodies.
Have a look at this Spot scene and that Spot model to get some ideas.
Cheers
Re: Transfer a locomotion policy of a quadruped robot from isaacsim
Hi~ Who created these scenes? Are these official scenes? Are there more demos like this?coppelia wrote: ↑30 Oct 2024, 09:44 Hello,
URDF portability across applications can be quite challenging. The URDF and your model are extremely heavy and not optimized. Also, the joint control mode should be set to force/torque in your case and many parameters adjusted, such as the armature parameter of the joints (MuJoCo). If running with other engines, you probably also have to artificially increase the inertia of most bodies.
Have a look at this Spot scene and that Spot model to get some ideas.
Cheers
Re: Transfer a locomotion policy of a quadruped robot from isaacsim
We have created those models/scenes.
Cheers
Cheers