华北平原暮色苍茫，太阳在落山。去年第一次在傍晚北上，静潭沉沉，渔船与船夫宛如静止，一度令我失语。今年再看，“长河落日圆”浑然天成，水的目的不是蜿蜒点缀，而是倒映天色，叠加玉璧一般的沉着，烁白的滩光残照反射，铺展些许晕开的绛紫。我只是拍照，再没有提笔。堂吉诃德战斗过的风车，粗壮的风车，在紫橘色的天际。云头是混色的梦，步步移动，太阳卡在风车的二叶之间。

——原来我提笔此刻路过的是泰安，我看见亮着灯的车站，高高的钢铁铁塔上亮着红红的大字，缚着亮红的霓虹：“宾馆”

我还只在感慨，在北京我极罕看天，天色美则美矣，不过如此。与广袤的华北平原相比，北京的天显得如此促狭，局促。所有的长河落日，至少要坐拥一片一眼望不到头的平原，水从脚下流向旷远，巨大的红日浑然安卧。

——天已经黑了，霓虹啊闷闷闪闪。

这坦坦荡荡，安坐如素的气势！云气啊，云彩啊，都是点缀罢了，真正的绚烂是那生生不息燃烧亘古的金乌！红得洋溢，红得热烈，一切都黯然失色，要么在凿壁偷光。而茶色的大地——它本身便没有太多的色彩，也不需借用阳光——稳重地承托着天空，红日，远处的炊烟，河流，一条公路横贯而过，上面的拖拉机在我看来也是静止的，也被大地载住。有的潭水冰封了，有的田地垦平了，乔木糊住我的视窗，一轮红日兮曩时当空。

我在车上正翻阅《禅与摩托车维修技术》，主角正骑着车在公路。那落日的情形，又和《在路上》的情形何等相似，“落日照在广袤的大平原上”。我如得神启，落日以其盛大，必须在大平原上演。“我想念迪恩·莫里亚蒂”凯鲁亚克写。而我想念我五小时之后的目的地。我在出发的时候，忽然想到“南望桂水，哭我故人”，这与我并不恰当，对我而言应该写作“南望弋江，思我故乡”。在我家乡，弋江自南向北，故为江东之地。我在江岸生活了十八年，竟不知道行于江上看落日是何等情致，而我在江滨总能看到对岸。是的，长江广阔，也能看到对岸。如此满江金辉碎鳞，与大平原迥然不同。“我忽然注意到，大地现在变得一片平坦，没有小丘，甚至也没有凸起之处。这表示我们已经进入红河谷，很快就会到达北达科他州了。”

“Ach, du lieber！”我正看到这一句话。

其实这一切尚未引发我写作的想法，直到车窗外出现山丘，红黯的天、影子似的山。小镇似乎有一层浮在上空的霓虹灯，我的意思是，霓虹的金色离地面有一段距离，我想是足够容纳人间的距离。小镇被山合围，也簇拥着中心一座顶高的山。

——“还有五分钟就要到了，妈妈！”

——“怎么又打电话？我在高铁上。”“高铁是不是在天上……？”

车厢是喧闹的，到滕州了，我打量这辆从北京出发的车，几乎没有上车的旅客，人们在不断地到家。

我开始臆想这样的小镇能激发什么样的故事，就如同阿加莎在东方快车上写出了东方快车谋杀案，想霓虹一样。我想到有银翼杀手在霓虹层之上滑动游走，飞檐走壁，我想象大山如大手向小镇缓缓靠近……然而我意识到，悲哀的是一切都已被写过，原来一切的故事在漫长的岁月里都有人写过，逆光的山峦只剩下黝黑的剪影，晚霞淋漓，无数悲欢发生过在这片土地。

Steps to install isaaclab & isaacsim

About nvidia driver,both from .run and from “Softwares & Updates” are available. But remember to choose open-kernal version if you come up with “No Devices Found”after “nvidia-smi”. If you use .run to install, you need to run “-m open-kernal” for fixing.

If you installed the wrong version, there will be “ppm init failed”. Use “alt +F2” to get into tty, then run “apt purge nvidia*” to uninstall all the driver related packages.

Download isaacsim to install isaaclab and you need to run “source ./setup_conda_env.sh” in the isaacsim folder every time you use the apps.

Steps to run Beyondmimic

We have errors like “KeyError: ‘obs_groups’ when running RSL-RL training”“ValueError: Observations for they key ‘rnd_state’ not found in infos[‘observations’].” “PPOConfig.init() got an unexpected keyword argument ‘model_name’ “ and other Keyerror.

--> Use rsl_rl v2.3.1 (if it’s too high will have Observations not found,if it’s too low will be PPOerror)

Steps to enable .xml into isaaclab

We use sim_utils.MjcfFileCfg to load .xml file.It is only accessed in main branch of Isaaclab. So I need switch to Isaaclab, create a new conda environment named “isaaclab2.3” in python3.11, use “pip” method to install isaacsim as https://isaac-sim.github.io/IsaacLab/main/source/setup/installation/pip_installation.html .Then run “./isaaclab.sh --install”

There are issues with....

With “ RPC failed; curl 56 GnuTLS recv error (-9) when running git clone”, which was network error.Solution: remenber to config the ssh-key.

git config --global http.postBuffer 1048576000 //Increase the Git buffer size
git config --global http.sslVerify false //Disable SSL verification temporarily
git config --global url."git@github.com:".insteadOf "https://github.com/"  //use SSH instead of HTTPS

And my disk has limited space, so I use “ls -n” to create a symbolic link from .conda to another disk.

Steps to process OMOMO

To install pytorch3d,run the scripts below:

conda create -n p3d python=3.9 -y
conda activate p3d
pip install torch==1.13.0+cu116 torchvision==0.14.0+cu116 \
  --extra-index-url https://download.pytorch.org/whl/cu116
# use pip not conda!!
pip install pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py39_cu116_pyt1130/download.html

Scipy can’t be installed directly in python 3.9:

python3.9 -m pip install scipy --user
pip3 install scipy --user

Install bps_torch:

pip install git+https://github.com/otaheri/chamfer_distance
pip install git+https://github.com/otaheri/bps_torch

And there still exist some problems with dimension arrays.repeat the 1 dimension data (.repeat(move_hand.shape[0]))to steptime’s and also remember the .long() method may cause data wrong.

The p3d environment is used to run python process_obj_scale.py to scale the objects.

Steps to train in Beyondmimic

If checkout to v2.2.1 , I can’t use smpl.py (MjcfFileCfg) to load model. If checkout to main, generate error “ModuleNotFoundError: No module named ‘isaacsim.asset’”

Choose to use Isaaclab in main, Isaacsim in 4.5.0

When use ./isaaclab.sh -i

• ModuleNotFoundError: No module named ‘pkg_resources’ -> the version of setuptools is too high

  • Install pip install flatdict==4.0.1 manually, repeat ./isaaclab.sh -i and pip install flatdict several times and you can dramatically solve the conflict problem.

  • Or use “conda install setuptools wheel packaging=23.0” to avoid flatdict version error.

• Network error : use ssh

    git config --global url."git@github.com:".insteadOf "https://github.com/"
  

Use ./isaaclab.sh -i none to replace the whole install

• Refer to https://blog.csdn.net/2303_76350015/article/details/155522236 install

    pip install "torch==2.7.0" "torchvision==0.22.0" \
       --index-url https://download.pytorch.org/whl/cu128
  

Change into Isaaclab in main, Isaacsim in 5.1.0

• For “ImportError: cannot import name ‘dump_pickle’ from ‘isaaclab.utils.io’”

  • Change train.py like this : https://github.com/BDX-R/BDX-R-IsaacLab/pull/14/changes

• For “doesn’t have a member called normalizer” when I want to play.py/--loginer wandb/train humanoid

  • Directly delete “normalizer=ppo_runner.obs_normalizer”

• isaaclab 0.54.3 requires starlette==0.49.1, but you have starlette 0.45.3 which is incompatible.

  • Use conda install “starlette==0.45.3”-> retry pip many times..

• Isaacsim is unable to record video

  • Install ffmpeg “conda install -c conda-forge ffmpeg”