CHENGJIE FANG / RESEARCH ARCHIVEUCAS 2026.11
← Field notes

Field note

Running DreamWaQ and Isaac Gym on an Ubuntu Server

Reproduction notes on Python compatibility, editable packages, NumPy, and headless training on a Tesla P4.

I was reproducing DreamWaQ on an Ubuntu server with a Tesla P4. This note records the dependency versions and errors I encountered in November 2024.

Use the Python version expected by Isaac Gym

My Python 3.10 environment failed immediately:

ERROR: Package 'isaacgym' requires a different Python: 3.10.14 not in '<3.9,>=3.6'

I created a Python 3.8 environment:

conda create -n isaacgym python=3.8
conda activate isaacgym

Useful environment-management commands are:

conda env list
conda deactivate
# Remove an environment only when it is no longer needed:
conda remove -n user_env --all

Install Isaac Gym and DreamWaQ

Download Isaac Gym Preview 4 from NVIDIA, then extract and install its Python package:

tar -xzvf IsaacGym_Preview_4_Package.tar.gz
pip install -e ./isaacgym/python/

This installation also installs the PyTorch dependency required by that package. Next, install DreamWaQ’s local libraries:

git clone https://github.com/Manaro-Alpha/DreamWaQ.git
cd DreamWaQ
pip install -e ./rsl_rl-1.0.2
pip install -e ./legged_gym

The editable installation completed with:

Installing collected packages: legged_gym
  Running setup.py develop for legged_gym
Successfully installed legged_gym-1.0.0

TensorBoard and NumPy

I also needed TensorBoard and a NumPy version compatible with this older stack:

pip install tensorboard
pip install numpy==1.23.5

The NumPy install replaced version 1.24.4. A separate pip uninstall numpy was unnecessary.

Start training and select a GPU

python3 ./legged_gym/legged_gym/scripts/train.py --task=go1

The original script also accepted explicit simulation and graphics devices:

python3 ./legged_gym/legged_gym/scripts/train.py \
  --task=go1 --graphics_device_id 1 --sim_device cuda:1

Alternatively, expose a selected physical GPU to the process:

CUDA_VISIBLE_DEVICES=1 python3 ./legged_gym/legged_gym/scripts/train.py \
  --task=go1 --headless

With CUDA_VISIBLE_DEVICES=1, that physical GPU is normally enumerated as cuda:0 inside the process; do not assume its physical index remains unchanged.

A segmentation fault in the graphics path

On the server, the run initialized PyTorch and PhysX, then crashed:

PyTorch version 1.11.0+cu113
Device count 1
Loading extension module gymtorch...
Setting seed: 5
Not connected to PVD
+++ Using GPU PhysX
Physics Engine: PhysX
Physics Device: cuda:0
GPU Pipeline: enabled
Segmentation fault (core dumped)

Running without a viewer solved this particular failure:

python3 ./legged_gym/legged_gym/scripts/train.py --task=go1 --headless

The training log then advanced normally. At iteration 90, my recorded output included:

Learning iteration 90/3000
Computation: 18159 steps/s (collection: 4.440s, learning: 0.973s)
Value function loss: 0.0032
Surrogate loss: -0.0007
Autoenc function loss: 5.9655
Mean action noise std: 0.56
Mean reward: 3.77
Mean episode length: 507.48
Total timesteps: 8945664
Iteration time: 5.41s
Total time: 429.71s

These are progress diagnostics from that run, not a final benchmark result. I also saved a driver-compatibility troubleshooting reference for installations on newer hardware.

Research figure