[Question] Implementing soft terrain for legged locomotion

[jnskkmhr]

My question is somewhat similar to this issue #90.

I am currently working on implementing a soft terrain contact model (more like a force model) based on resistive force theory described in this paper: https://arxiv.org/abs/1303.7065.

As an initial test, I dropped a robot foot from a height of 0.5m and observed that the terrain can support the object with a certain level of penetration.

My question is: what could be a promising approach to simulate this soft terrain model for legged robot locomotion, particularly for bipedal locomotion?
I found that simply applying force to the foot does not work effectively due to the resulting moments causing the robot to flip.
Do you think increasing contact points will help?
I already tried with 5 points (one at the center, 4 at the corner), but this did not work.

I can share my current code, which essentially computes the contact force based on the paper's formulation and applies it to the robot using the articulation API.

rft.mp4

[RandomOakForest]

Thank you for posting this. The best solution we found for this, currently, is that you can use WARP to model a soft terrain. There is an example script here that may be of help. Also, in our showroom page, there is some information about procedurally generated terrains with different configurations.

[jnskkmhr]

Hi @RandomOakForest
Thank you for the response.

I thought about that too, but I did not understand how to reflect the physics done in warp to IsaacLab simulation (how to integrate robot-particle interaction done by warp to Isaac?)
Also, as far as I tried both warp and PhysX particle system, running particle sim is really expensive.

Is there a way to apply force computed in warp to every vertices of mesh (like foot)?
It would be great if I can run sort of like reduced order terrain model (kind of like model which approximate foot-terrain interaction).
Then apply forces computed by this terrain model to foot mesh, that is my idea.

[RandomOakForest]

Thanks for following up. I'll move this into our Discussions section for the team to follow up.

[jnskkmhr]

Hi, @RandomOakForest I forgot to follow up on this.
It seems my initial implementation is somewhat functional.
Note that there is no collider activated in this simulation, so the contact is computed based on my custom model.

Right now, it has much room to improve though.
For example, I see horizontal drift behavior when robot's command velocity gets higher.

It would be great if Nvidia team is interested in this direction.

rft.mp4

[Asthestarsfalll]

Hi do you fix the drift?

[jnskkmhr]

3d_rft_G1_locomotion_result.mp4

[Asthestarsfalll]

wow, this is really cool.
When you release code, I’d appreciate it if you could remind me!

[istarendil]

Hi @jnskkmhr,

Thanks for sharing your progress, this approach is very interesting. Do you have any news about this?
I am also trying to implement 3d-RFT for training model in Isaaclab.

[jnskkmhr]

@istarendil
Hi, I am trying to wrap up journal submission for this.
Code will be released once I put paper in arxiv.
But, here is preliminary results for ICRA workshop: https://drive.google.com/file/d/1vk3Zvnw2LrQx_qtvTPM4cS9aPcHFdFbl/view?usp=drive_link

[istarendil]

@jnskkmhr,
Great results, I will be very glad to also read the paper.