Integration of HumanoidBench tasks#339
Integration of HumanoidBench tasks#339MoritzMeser wants to merge 3 commits intogoogle-deepmind:mainfrom
Conversation
|
Great tasks! It looks like the floor is visualized a bit higher than the contact plane, let's modify the floor so that contacts with it are properly visualized. For the H1 push task, can we make the box goal position interactive? We can use the same logic as the quadruped flat task to enable the user to move the goal mocap body via click and drag. For the G1 walking task, have you tried using the same PD actuator model as the H1? Additionally, we might want to remove the hand DOFs from the model since they aren't used for locomotion and the additional DOFs make simulation more expensive. We should be able to reduce the planning horizon for this task to 0.35-0.4 seconds. Another way to improve this task is to replace the mesh collision geometries with primitives (eg, spheres, capsules). Thanks! |
|
Yes, sure. I'll work through the points and then get back to you. |
…k for the Unitree H1 robot readme
…teractive and removed the G1 robot
|
Hi I solved the issue with the floor and made the box goal interactive. Regarding the G1 robot, I have remove the DOFs in the hand, but MPC still does not manage to keep standing. So I removed the G1 completely. |
Hello.
This is a new pull request to integrate some of the tasks from HumanoidBench into MJPC.
As discussed in the previous pull request (#328, which is now closed) and by mail, I have solved the issues we discussed earlier.
I updated the git tag for Menagerie to a version that includes the Unitree G1 robot. Because of this, I also had to update the patch file for the quadruped robot, so it is still compatible. I just copied the patch file there, which is on the "deepmind" branch of this repository.