search

lnotspotl / a1_sim_py

This is a temporary repository, which contains all the files needed to simulate the Unitree Robotics' a1 quadrupedal robot.
Mozilla Public License 2.0
65 stars 29 forks source link

Do you how to get observation as same as issacgym's obervation ? #5

Open heartInsert opened 1 week ago

heartInsert commented 1 week ago

hey , bro, I trained a RL model in issacgym , and I run this al_sim_py sucessfully , there is a observation called dof_state in issacgym , maybe it is called joint_state in gazebo , how can I get this joint_state ?

lnotspotl commented 1 week ago

You should be able to get the joint angles from the "a1_gazebo/joint_states" topic, see here

By the way, I am curious, what kind of a policy have you trained?

heartInsert commented 1 week ago

Thanks for your relpy , bro . After I subscribe the topic a1_gazebo/joint_states , I get WARNING: no messages received and simulated time is active. Is /clock being published? . I don't know why it happend. So I use service /gazebo/get_joint_properties 12 times to get my joint state .

heartInsert commented 1 week ago

I have trained a PPO model in Isaacgym and I have to deploy this PPO model in A1 ， the biggest problem right now is how to change gazebo's observation to Isaacgym's observation . I have accomplish the whole workflow ,
my model walks well in Isaacgym but the doesn't walk well in gazebo . I try to figure it out.

lnotspotl commented 1 week ago

It's paramount your model is fed the same input it was given during the training phase. It's a bit tricky, I am talking from my own experience. Take a look at the tbai project where I was trying to deploy a trained walking policy in Gazebo, exactly what you are trying to achieve.

heartInsert commented 5 days ago

Thanks for your reply , bro. After carefully reading a lot of different projects , I finally overcome this problem , my PPO model could walks in Gazebo right now . Thanks for your help again.

lnotspotl commented 5 days ago

That's awesome. Congratulations