Hello Nnamdi, I hope you are doing well.
I want to share this repo (https://github.com/uwsbel/sbel-reproducibility/tree/master/2025/multi-terrain-RL) with you, which does the RL training for the demo videos I shared earlier. The specific robot type or task might differ, but I think the structure or idea would be similar. I realize I previously offered to meet, but after giving it more thought—and in light of similar requests from others—We’ve decided it's better to share a prepared code repo instead. This way, you can try it on your own and post any questions here, so others can also benefit from the discussion. Best, Harry On Wednesday, July 30, 2025 at 7:50:49 PM UTC-5 [email protected] wrote: > Hello Harry, > > Your simulation results looks great and are similar to what I hope to > implement in my setup. I am interested in discussing more about this and > would love to set up a meeting as soon as possible. Can you please let me > know what day/time you are available? > > Best regards, > Nnamdi > > On Wednesday, July 30, 2025 at 1:50:56 PM UTC-4 [email protected] wrote: > >> Picture attachment for last message: >> [image: Screenshot from 2025-07-30 10-45-41.png] >> Best, >> Harry >> On Wednesday, July 30, 2025 at 11:17:42 AM UTC-5 Harry ZHANG wrote: >> >>> Hello Nnamdi, >>> >>> My name's Harry and Dan asked me to follow up on this thread. >>> >>> About couple weeks ago, we start exploring using PyChrono as our physics >>> engine to train RL policies for quadruped locomotion task. Our projects's >>> goal is to investigate simulation terrain's effect on locomotion policy >>> performance. In PyChrono there are different fidelity levels for terrain >>> model (from low to high): rigid terrain (rtf<<1) --> SCM deformable terrain >>> (rtf~=1)-->CRM granular terrain(rtf~=5), where rtf means real time fator. >>> RL library we are using is "RSL-RL". >>> >>> Please check this link for prelim results we got: >>> https://drive.google.com/drive/folders/1Yhx1goYDsv0FZdk4OfyH9_yRazkRSwMQ?usp=sharing. >>> >>> The locomotion policy trained for 0.5 m/s longitudinal speed walking. The >>> policy trained with rigid terrain (about 6 hours) and we put it on scm >>> deformable terrain for testing purpose, in both rigid and deformable >>> terrain the evaluation looks decent. >>> >>> For us, there are two major things in our todo list: 1) implement some >>> parallelzation methods during training. current 6 hour training time is >>> trained with one single simulation agent interact with RSL-RL algorithm, we >>> try to see if it's possible to do faster by training with multiple >>> simulation agents. 2) Include quadrupeds with CRM terrain in both policy >>> training and evaluation. See the attched picture, we have a simulation >>> script includes quadrupeds on CRM terrain, but we need to merge this into >>> RL pipeline. >>> >>> Let me know if you have any questions, I am available for meeting if you >>> want me explain more about the pipeline and details. >>> >>> Best, >>> Harry >>> >>> On Monday, July 28, 2025 at 4:40:48 PM UTC-5 [email protected] wrote: >>> >>>> Thank you so much for your response. >>>> >>>> We are starting to work in this area as well, from a slightly different >>>> perspective. Harry, a great student in the lab, will share his findings in >>>> about a week, to steer you in the right direction. Are you working with a >>>> quadruped or something similar? >>>> >>>> I am working on a small-sized bioinspired quadruped robot, and I look >>>> forward to the findings from your student. >>>> >>>> Is it like (a) the entire terrain is made up of rocks 2cm to 6 cm, or >>>> (b) you have some deformable terrain that once in a while has some rocks 2 >>>> cm to 6 cm scattered around? This is important, since (a) will likely take >>>> more time to run than (b). >>>> >>>> I am interested in having the entire terrain composed of movable rocks, >>>> which the limbs can interact with, such as each terrain box filled with >>>> rocks of a specific size (2cm to 6cm). I am also open to suggestions on >>>> whether this setup might be computationally intensive and any >>>> alternatives, >>>> such as option B. The image below is what I have in mind (generated in >>>> MuJoCo, but is fixed to the ground) >>>> [image: Screenshot 2025-07-28 172945.png] >>>> >>>> In principle, this should be doable but we haven’t tried it yet. The >>>> needed modeling elements are in Chrono, it’s just a matter of putting a >>>> model like this together and seeing how it does. Easier said than done, >>>> but >>>> we might be able to provide some feedback when you cross that bridge. >>>> >>>> Thank you. I will provide more details on my proposed implementation >>>> soon. >>>> >>>> In about one week we’ll share a way to install a newer version of >>>> PyChrono, a beta version, that has support for what’s called in Chrono >>>> “the >>>> CRM terrain” model. Might be helpful for you. Needs an NVIDIA GPU to run. >>>> >>>> Looking forward to the new version. I have been facing some SDL2 >>>> package errors with my current version. I gave more details in a different >>>> thread. >>>> On Thursday, July 24, 2025 at 3:53:14 PM UTC-4 Dan Negrut wrote: >>>> >>>>> Please see my comments in this font below. >>>>> >>>>> >>>>> >>>>> Thank you, >>>>> >>>>> Dan >>>>> >>>>> --------------------------------------------- >>>>> >>>>> Bernard A. and Frances M. Weideman Professor >>>>> >>>>> NVIDIA CUDA Fellow >>>>> >>>>> Department of Mechanical Engineering >>>>> >>>>> Department of Computer Science >>>>> >>>>> University of Wisconsin - Madison >>>>> >>>>> 4150ME, 1513 University Avenue >>>>> >>>>> Madison, WI 53706-1572 >>>>> >>>>> 608 772 0914 <(608)%20772-0914> >>>>> >>>>> http://sbel.wisc.edu/ >>>>> >>>>> http://projectchrono.org/ >>>>> >>>>> --------------------------------------------- >>>>> >>>>> >>>>> >>>>> *From:* Nnamdi Chikere <[email protected]> >>>>> *Sent:* Thursday, July 24, 2025 10:31 AM >>>>> *To:* Dan Negrut <[email protected]> >>>>> *Cc:* ProjectChrono <[email protected]> >>>>> *Subject:* Re: [chrono] Help with Granular Simulation on Windows >>>>> using PyChrono >>>>> >>>>> >>>>> >>>>> Hello Professor, >>>>> >>>>> >>>>> >>>>> Thank you so much for your reply. >>>>> >>>>> >>>>> >>>>> For the first stage, I am trying to optimize the parameters for >>>>> control of the robot’s locomotion on complex environments including sand >>>>> and rocks using CPGs. The end goal is to later use RL to implement >>>>> adaptive >>>>> control with changing terrains. >>>>> >>>>> We are starting to work in this area as well, from a slightly >>>>> different perspective. Harry, a great student in the lab, will share his >>>>> findings in about a week, to steer you in the right direction. Are you >>>>> working with a quadruped or something similar? >>>>> >>>>> >>>>> >>>>> For the rock sizes, I was looking to implement three rock sizes >>>>> ranging between 2cm and 6cm filled in a box on which the robot will >>>>> operate. >>>>> >>>>> Is it like (a) the entire terrain is made up of rocks 2cm to 6 cm; or >>>>> (b) you have some deformable terrain that once in a while has some rocks >>>>> 2 >>>>> cm to 6 cm scattered around? This is important, since (a) will likely >>>>> take >>>>> more time to run than (b). >>>>> >>>>> >>>>> >>>>> I am also trying to implement a flexible limb that will be attached to >>>>> the motors and deform when in contact with some force or the ground, is >>>>> it >>>>> possible to implement this to a specified link? >>>>> >>>>> In principle, this should be doable but we haven’t tried it yet. The >>>>> needed modeling elements are in Chrono, it’s just a matter of putting a >>>>> model like this together and seeing how it does. Easier said than done, >>>>> but >>>>> we might be able to provide some feedback when you cross that bridge. >>>>> >>>>> >>>>> >>>>> I am working with Python and windows, and would appreciate some help >>>>> as regards this >>>>> >>>>> In about one week we’ll share a way to install a newer version of >>>>> PyChrono, a beta version, that has support for what’s called in Chrono >>>>> “the >>>>> CRM terrain” model. Might be helpful for you. Needs an NVIDIA GPU to run. >>>>> >>>>> >>>>> >>>>> Best regards, >>>>> >>>>> Nnamdi >>>>> >>>>> >>>>> >>>>> On Thu, Jul 24, 2025 at 2:10 AM Dan Negrut <[email protected]> wrote: >>>>> >>>>> If you describe what you need to do with your robot, we might be able >>>>> to steer you in the right direction. >>>>> How big are the rocks you mentioned? >>>>> >>>>> What sort of end-goal do you have? Design of robot, control of robot, >>>>> RL, optimal design, etc. >>>>> >>>>> >>>>> >>>>> Dan >>>>> >>>>> --------------------------------------------- >>>>> >>>>> Bernard A. and Frances M. Weideman Professor >>>>> >>>>> NVIDIA CUDA Fellow >>>>> >>>>> Department of Mechanical Engineering >>>>> >>>>> Department of Computer Science >>>>> >>>>> University of Wisconsin - Madison >>>>> >>>>> 4150ME, 1513 University Avenue >>>>> <https://urldefense.com/v3/__https:/www.google.com/maps/search/1513*University*Avenue**A0D*0A*Madison,*WI*53706-1572?entry=gmail&source=g__;KysrJSUrKys!!Mak6IKo!Od8BEUkPpm5WoEens9KtQIg47iogD1i2m9LDYGHeEv2M8GiWuVl9N0Stk5HjyF-DIbSFX75lMxV4hhD5$> >>>>> >>>>> Madison, WI 53706-1572 >>>>> <https://urldefense.com/v3/__https:/www.google.com/maps/search/1513*University*Avenue**A0D*0A*Madison,*WI*53706-1572?entry=gmail&source=g__;KysrJSUrKys!!Mak6IKo!Od8BEUkPpm5WoEens9KtQIg47iogD1i2m9LDYGHeEv2M8GiWuVl9N0Stk5HjyF-DIbSFX75lMxV4hhD5$> >>>>> >>>>> 608 772 0914 <(608)%20772-0914> >>>>> >>>>> http://sbel.wisc.edu/ >>>>> >>>>> http://projectchrono.org/ >>>>> <https://urldefense.com/v3/__http:/projectchrono.org/__;!!Mak6IKo!Od8BEUkPpm5WoEens9KtQIg47iogD1i2m9LDYGHeEv2M8GiWuVl9N0Stk5HjyF-DIbSFX75lM0FtMmwz$> >>>>> >>>>> >>>>> --------------------------------------------- >>>>> >>>>> >>>>> >>>>> *From:* 'Nnamdi Chikere' via ProjectChrono <[email protected]> >>>>> >>>>> *Sent:* Tuesday, July 22, 2025 6:16 PM >>>>> *To:* ProjectChrono <[email protected]> >>>>> *Subject:* [chrono] Help with Granular Simulation on Windows using >>>>> PyChrono >>>>> >>>>> >>>>> >>>>> Hello. >>>>> >>>>> I am working on a simulation where a robot moves on different >>>>> terrains, and I want to simulate the robot moving on sand and rocks. I >>>>> have >>>>> been able to implement the soil terrain using the SCM parameters in >>>>> PyChrono. Is it possible to implement granular media, such as rocks, in >>>>> PyChrono? I attempted to install the DEME package, but encountered >>>>> several >>>>> errors. I am working with PyChrono 9.0 and Python 3.9. >>>>> >>>>> Thank you. >>>>> >>>>> -- >>>>> You received this message because you are subscribed to the Google >>>>> Groups "ProjectChrono" group. >>>>> To unsubscribe from this group and stop receiving emails from it, send >>>>> an email to [email protected]. >>>>> To view this discussion visit >>>>> https://groups.google.com/d/msgid/projectchrono/d38be8a1-f6e0-4661-83ba-d51b500ccea5n%40googlegroups.com >>>>> >>>>> <https://urldefense.com/v3/__https:/groups.google.com/d/msgid/projectchrono/d38be8a1-f6e0-4661-83ba-d51b500ccea5n*40googlegroups.com?utm_medium=email&utm_source=footer__;JQ!!Mak6IKo!IHl25Nn94rVvk-8cbLYWCsyZEpu7zcoKd_JY1u7_ZFpAebs2_ga8vymx9iyQT1Ywjzy8INBVzHAHobmHZEkRFozUi6s$> >>>>> . >>>>> >>>>> -- You received this message because you are subscribed to the Google Groups "ProjectChrono" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/projectchrono/063cba57-eb2a-44d2-a3b8-e5fd521bec87n%40googlegroups.com.
