Isaac Gym Segmentation Fault

Description: Segmentation fault caught when launching Isaac Gym inside docker container
The reason of segmentation fault is the container could not use the RTX graphic card to render led to segmentation fault when loading Isaac Gym’s viewer
Inspected log:

Creating 36 environments
Animating DOF 0 ('abdomen_z')
Unhandled descriptor set 433
Unhandled descriptor set 414286096
Unhandled descriptor set 431466320
Segmentation fault (core dumped)

Inside the container, inspect the renderer

vulkaninfo --json |grep "deviceName"

# Output:
deviceName : llvmpipe (LLVM 12.0.0, 256 bits)   ⟵ software renderer

llvmpipe means Vulkan has fallen back to CPU rasterization because it cannot find a valid NVIDIA driver.

The container cannot see NVIDIA’s Vulkan ICD manifest (nvidia_icd.json) and/or the matching driver libraries, so the Vulkan loader skips the GPU and uses the software fallback; Isaac Gym crashes when it tries to create GPU-only resources.

Cause: Incorrect mount command is -v /usr/lib/x86_64-linux-gnu/libnvidia-gl-550.so.0), see the following lines:

#!/bin/bash
set -e
set -u

# --- IMPORTANT ---
# Replace :0 with the value you found in Step 1
export DISPLAY=:1

echo "Setting display to $DISPLAY"

# Temporarily allow local connections from Docker to the X server
xhost +local:

# Run the Docker container
docker run -it --rm \
  --gpus all \
  -e NVIDIA_DRIVER_CAPABILITIES=all \
  -e DISPLAY=$DISPLAY \
  -v /tmp/.X11-unix:/tmp/.X11-unix \
  -v /usr/lib/x86_64-linux-gnu/libnvidia-gl-550.so.0:/usr/lib/x86_64-linux-gnu/libnvidia-gl-550.so.0:ro \ <----- HERE
  -v /usr/lib/x86_64-linux-gnu/libGLX_nvidia.so.0:/usr/lib/x86_64-linux-gnu/libGLX_nvidia.so.0:ro \
  -v .:/agentic_robot_assembly \
  --network=host \
  --name isaacgym_container \
  isaacgym:py38-cuda \
  /bin/bash

# Revert xhost settings
echo "Restoring display access control"
xhost -local:

It was mounted incorrect nvidia driver version 550, this would fail if the host machine is using Nvidia driver version orther 550.

Solution: Mount only the Vulkan ICD Here is the updated run.sh:

#!/usr/bin/env bash
set -euo pipefail

: "${DISPLAY:=:1}"
export DISPLAY
echo "[run.sh] DISPLAY=$DISPLAY"

###############################################################################
if   [ -f /usr/share/vulkan/icd.d/nvidia_icd.json ]; then
      ICD_DIR=/usr/share/vulkan/icd.d
elif [ -f /etc/vulkan/icd.d/nvidia_icd.json ]; then
      ICD_DIR=/etc/vulkan/icd.d
else
      echo "[run.sh] ERROR: nvidia_icd.json not found on host." >&2
      exit 1
fi
echo "[run.sh] Using Vulkan ICD dir $ICD_DIR"

VOLS=(-v /tmp/.X11-unix:/tmp/.X11-unix          # X socket
      -v "$ICD_DIR:$ICD_DIR:ro"                # Vulkan ICD JSON
      -v "$(pwd)":/agentic_robot_assembly)     # 專案

ENV_VARS=(-e DISPLAY="$DISPLAY"
          -e VK_ICD_FILENAMES="$ICD_DIR/nvidia_icd.json")

xhost +local: 1>/dev/null

docker run -it --rm \
  --gpus all \
  -e NVIDIA_DRIVER_CAPABILITIES=all,graphics,display,utility,compute \
  "${ENV_VARS[@]}" \
  "${VOLS[@]}" \
  --network host \
  --name isaacgym_container \
  isaacgym:py38-cuda \
  /bin/bash

xhost -local: 1>/dev/null