diff --git a/.github/workflows/ubuntu-wheel-sbsa.yml b/.github/workflows/ubuntu-wheel-sbsa.yml
new file mode 100644
index 00000000000..54a3cf5e89f
--- /dev/null
+++ b/.github/workflows/ubuntu-wheel-sbsa.yml
@@ -0,0 +1,194 @@
+name: Ubuntu Wheel SBSA CUDA
+permissions: {}
+
+on:
+  workflow_dispatch:
+    inputs:
+      developer_build:
+        description: 'Set to OFF for Release wheels'
+        required: false
+        default: 'ON'
+  push:
+    branches:
+      - main
+  pull_request:
+    types: [opened, reopened, synchronize]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+env:
+  GCE_CLI_GHA_VERSION: '416.0.0'      # Fixed to avoid dependency on API changes
+  BUILD_CUDA_MODULE: 'ON'
+
+jobs:
+  build-wheel:
+    permissions:
+      contents: write  # Release upload
+    name: Build SBSA CUDA wheel
+    runs-on: ubuntu-24.04-arm
+    strategy:
+      fail-fast: false
+      matrix:
+        python_version: ['3.10', '3.11', '3.12', '3.13', '3.14']
+        is_main:
+          - ${{ github.ref == 'refs/heads/main' }}
+        exclude:
+          - is_main: false
+            python_version: '3.10'
+          - is_main: false
+            python_version: '3.11'
+          - is_main: false
+            python_version: '3.12'
+          - is_main: false
+            python_version: '3.13'
+    env:
+      DEVELOPER_BUILD: ${{ github.event.inputs.developer_build || 'ON' }}
+      PYTHON_VERSION: ${{ matrix.python_version }}
+      CCACHE_TAR_NAME: open3d-ubuntu-2204-cuda-sbsa-ci-ccache
+      OPEN3D_CPU_RENDERING: true
+      BUILD_PYTORCH_OPS: 'ON'
+      # TensorFlow v2.20 only supports python 3.9-3.13
+      BUILD_TENSORFLOW_OPS: ${{ matrix.python_version >= '3.14' && 'OFF' || 'ON' }}
+    steps:
+      - name: Checkout source code
+        uses: actions/checkout@v6
+      - name: Maximize build space
+        run: |
+          source util/ci_utils.sh
+          maximize_ubuntu_github_actions_build_space
+      # Be verbose and explicit here such that a developer can directly copy the
+      # `docker/docker_build.sh xxx` command to execute locally.
+      - name: Docker build
+        run: |
+          if [ "${{ env.PYTHON_VERSION }}" = "3.10" ] && [ "${{ env.DEVELOPER_BUILD }}" = "ON"  ]; then
+            docker/docker_build.sh cuda_wheel_py310_dev
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.11" ] && [ "${{ env.DEVELOPER_BUILD }}" = "ON"  ]; then
+            docker/docker_build.sh cuda_wheel_py311_dev
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.12" ] && [ "${{ env.DEVELOPER_BUILD }}" = "ON"  ]; then
+            docker/docker_build.sh cuda_wheel_py312_dev
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.13" ] && [ "${{ env.DEVELOPER_BUILD }}" = "ON"  ]; then
+            docker/docker_build.sh cuda_wheel_py313_dev
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.14" ] && [ "${{ env.DEVELOPER_BUILD }}" = "ON"  ]; then
+            docker/docker_build.sh cuda_wheel_py314_dev
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.10" ] && [ "${{ env.DEVELOPER_BUILD }}" = "OFF" ]; then
+            docker/docker_build.sh cuda_wheel_py310
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.11" ] && [ "${{ env.DEVELOPER_BUILD }}" = "OFF" ]; then
+            docker/docker_build.sh cuda_wheel_py311
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.12" ] && [ "${{ env.DEVELOPER_BUILD }}" = "OFF" ]; then
+            docker/docker_build.sh cuda_wheel_py312
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.13" ] && [ "${{ env.DEVELOPER_BUILD }}" = "OFF" ]; then
+            docker/docker_build.sh cuda_wheel_py313
+          elif [ "${{ env.PYTHON_VERSION }}" = "3.14" ] && [ "${{ env.DEVELOPER_BUILD }}" = "OFF" ]; then
+            docker/docker_build.sh cuda_wheel_py314
+          fi
+          shopt -s failglob
+          PIP_PKG_NAME="$(basename ${GITHUB_WORKSPACE}/open3d-[0-9]*.whl)"
+          PIP_CPU_PKG_NAME="$(basename ${GITHUB_WORKSPACE}/open3d_cpu*.whl)"
+          echo "PIP_PKG_NAME=$PIP_PKG_NAME" >> $GITHUB_ENV
+          echo "PIP_CPU_PKG_NAME=$PIP_CPU_PKG_NAME" >> $GITHUB_ENV
+      - name: Upload wheel to GitHub artifacts
+        uses: actions/upload-artifact@v7
+        with:
+          name: ${{ env.PIP_PKG_NAME }}
+          path: |
+            ${{ env.PIP_PKG_NAME }}
+            ${{ env.PIP_CPU_PKG_NAME }}
+          if-no-files-found: error
+      - name: Update devel release
+        if: ${{ github.ref == 'refs/heads/main' }}
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          .github/workflows/update_release.sh ${GITHUB_WORKSPACE}/${{ env.PIP_PKG_NAME }} \
+          ${GITHUB_WORKSPACE}/${{ env.PIP_CPU_PKG_NAME }}
+      - name: GCloud CLI auth
+        if: ${{ github.ref == 'refs/heads/main' }}
+        uses: 'google-github-actions/auth@v3'
+        with:
+          project_id: ${{ secrets.GCE_PROJECT }}
+          credentials_json: '${{ secrets.GCE_SA_KEY_GPU_CI }}'
+      - name: GCloud CLI setup
+        if: ${{ github.ref == 'refs/heads/main' }}
+        uses: google-github-actions/setup-gcloud@v3
+        with:
+          version: ${{ env.GCE_CLI_GHA_VERSION }}
+          project_id: ${{ secrets.GCE_PROJECT }}
+      - name: Upload ccache to GCS
+        if: ${{ github.ref == 'refs/heads/main' }}
+        run: |
+          gsutil cp ${GITHUB_WORKSPACE}/${{ env.CCACHE_TAR_NAME }}.tar.xz gs://open3d-ci-cache/
+
+  test-wheel-cpu:
+    name: Test SBSA CUDA wheel CPU
+    permissions:
+      contents: read
+    runs-on: ubuntu-24.04-arm
+    needs: [build-wheel]
+    strategy:
+      fail-fast: false
+      matrix:
+        python_version: ['3.10', '3.11', '3.12', '3.13', '3.14']
+        is_main:
+          - ${{ github.ref == 'refs/heads/main' }}
+        exclude:
+          - is_main: false
+            python_version: '3.10'
+          - is_main: false
+            python_version: '3.11'
+          - is_main: false
+            python_version: '3.12'
+          - is_main: false
+            python_version: '3.13'
+    env:
+      OPEN3D_ML_ROOT: ${{ github.workspace }}/Open3D-ML
+      BUILD_PYTORCH_OPS: 'ON'
+      # TensorFlow v2.20 only supports python 3.9-3.13
+      BUILD_TENSORFLOW_OPS: ${{ matrix.python_version >= '3.14' && 'OFF' || 'ON' }}
+    steps:
+      - name: Checkout source code
+        uses: actions/checkout@v6
+      - name: Maximize build space
+        run: |
+          source util/ci_utils.sh
+          maximize_ubuntu_github_actions_build_space
+      - name: Checkout Open3D-ML source code
+        uses: actions/checkout@v6
+        with:
+          repository: isl-org/Open3D-ML
+          ref: main
+          path: ${{ env.OPEN3D_ML_ROOT }}
+      - name: Download wheels
+        uses: actions/download-artifact@v8
+        with:
+          pattern: open3d*-manylinux*.whl
+          merge-multiple: true
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: ${{ matrix.python_version }}
+      - name: Test Python package
+        run: |
+          python -V
+          source util/ci_utils.sh
+          pi_tag=$(python -c "import sys; print(f'cp{sys.version_info.major}{sys.version_info.minor}')")
+          test_wheel open3d-[0-9]*-"$pi_tag"-*.whl
+      - name: Run Python unit tests
+        run: |
+          source util/ci_utils.sh
+          echo "Running Open3D python tests..."
+          run_python_tests
+      - name: Test Python package (CPU)
+        env:
+          BUILD_CUDA_MODULE: OFF
+        run: |
+          python -V
+          source util/ci_utils.sh
+          pi_tag=$(python -c "import sys; print(f'cp{sys.version_info.major}{sys.version_info.minor}')")
+          test_wheel open3d_cpu*-"$pi_tag"-*.whl
+      - name: Run Python unit tests (CPU)
+        run: |
+          source util/ci_utils.sh
+          echo "Running Open3D python tests (CPU wheel)..."
+          run_python_tests
diff --git a/.github/workflows/windows.yml b/.github/workflows/windows.yml
index a375bea818b..c1e9eb85971 100644
--- a/.github/workflows/windows.yml
+++ b/.github/workflows/windows.yml
@@ -27,7 +27,7 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  PIP_VER: "24.3.1"
+  PIP_VER: "26.0.1"
   CUDA_VERSION: "12.6.0"
   SRC_DIR: "D:\\a\\open3d\\open3d"
   BUILD_DIR: "C:\\Open3D\\build"
@@ -76,7 +76,7 @@ jobs:
           # Installer arguments
           $CUDA_INSTALL_ARGS = "-s"
           # Required packages
-          $CUDA_PACKAGES = "nvcc", "visual_studio_integration", "cublas", "cublas_dev", "cudart", "cusolver", "cusolver_dev", "cusparse", "cusparse_dev", "npp", "npp_dev", "nvtx", "thrust"
+          $CUDA_PACKAGES = "nvcc", "crt", "visual_studio_integration", "cublas", "cublas_dev", "cudart", "cusolver", "cusolver_dev", "cusparse", "cusparse_dev", "npp", "npp_dev", "nvtx", "thrust"
           $CUDA_PACKAGES.ForEach({ $CUDA_INSTALL_ARGS += " $($_)_$($CUDA_VER)" })
           # Download and install CUDA
           echo "Downloading CUDA installer from $CUDA_URL"
@@ -112,6 +112,10 @@ jobs:
           if ( '${{ matrix.BUILD_CUDA_MODULE }}' -eq 'ON' ) {
             # python -m pip install -r open3d_ml/requirements-torch-cuda.txt
             python -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
+            if ($LASTEXITCODE -ne 0) {
+              Write-Host "Stable PyTorch with CUDA 12.6 not found, trying nightly..."
+              python -m pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu132
+            }
           } else {
             python -m pip install -r open3d_ml/requirements-torch.txt
           }
diff --git a/CHANGELOG.md b/CHANGELOG.md
index cdf555b0d11..35eef6b54be 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -74,6 +74,10 @@
 -   macOS x86_64 not longer supported, only macOS arm64 is supported.
 -   Python 3.13+3.14 support
 -   Fix Windows build failure for PyTorch ops due to PyTorch's bundled fmt (v11+) requiring `/utf-8` with MSVC (PR #7447)
+-   Upgrade CUDA to 12.6 and PyTorch to 2.12
+-   Add SBSA (ARM64) CUDA wheel builds in CI
+-   Refactor CI release upload steps into a reusable `update_release.sh` script
+-   Fix style workflow to push formatting commits to the correct PR branch
 
 
 ## 0.13
diff --git a/docker/Dockerfile.wheel b/docker/Dockerfile.wheel
index afbc45be5cb..f471a78abd4 100644
--- a/docker/Dockerfile.wheel
+++ b/docker/Dockerfile.wheel
@@ -1,5 +1,5 @@
 # FROM must be called before other ARGS except for ARG BASE_IMAGE
-ARG BASE_IMAGE=nvidia/cuda:12.6.3-cudnn-devel-ubuntu22.04
+ARG BASE_IMAGE=nvidia/cuda:12.6.0-cudnn-devel-ubuntu22.04
 FROM ${BASE_IMAGE}
 
 # Customizable build arguments from cuda.yml
@@ -7,6 +7,7 @@ ARG DEVELOPER_BUILD
 ARG CCACHE_TAR_NAME
 ARG CMAKE_VERSION
 ARG PYTHON_VERSION
+ARG BUILD_CUDA_MODULE=OFF
 ARG BUILD_TENSORFLOW_OPS
 ARG BUILD_PYTORCH_OPS
 ARG CI
@@ -17,6 +18,7 @@ ENV DEVELOPER_BUILD=${DEVELOPER_BUILD}
 ENV CCACHE_TAR_NAME=${CCACHE_TAR_NAME}
 ENV CMAKE_VERSION=${CMAKE_VERSION}
 ENV PYTHON_VERSION=${PYTHON_VERSION}
+ENV BUILD_CUDA_MODULE=${BUILD_CUDA_MODULE}
 ENV BUILD_PYTORCH_OPS=${BUILD_PYTORCH_OPS}
 ENV BUILD_TENSORFLOW_OPS=${BUILD_TENSORFLOW_OPS}
 
@@ -27,14 +29,16 @@ ENV SUDO=command
 
 SHELL ["/bin/bash", "-c"]
 
-# Fix Nvidia repo key rotation issue
+# Fix Nvidia repo key rotation issue (x86_64 only — ARM CUDA images are unaffected)
 # https://forums.developer.nvidia.com/t/notice-cuda-linux-repository-key-rotation/212771
 # https://forums.developer.nvidia.com/t/18-04-cuda-docker-image-is-broken/212892/10
 # https://code.visualstudio.com/remote/advancedcontainers/reduce-docker-warnings#:~:text=Warning%3A%20apt%2Dkey%20output%20should,not%20running%20from%20a%20terminal.
 RUN export APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn \
- && apt-key del 7fa2af80 \
- && apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/3bf863cc.pub \
- && apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/7fa2af80.pub
+ && if [ "$(uname -m)" = "x86_64" ]; then \
+      apt-key del 7fa2af80 \
+      && apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/3bf863cc.pub \
+      && apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/7fa2af80.pub; \
+    fi
 
 # Dependencies: basic
 RUN apt-get update && apt-get install -y \
diff --git a/docker/docker_build.sh b/docker/docker_build.sh
index 3eb2c188c3a..be5a2fd7e61 100755
--- a/docker/docker_build.sh
+++ b/docker/docker_build.sh
@@ -62,7 +62,9 @@ OPTION:
     3-ml-shared-jammy         : CUDA CI, 3-ml-shared-jammy (cxx11_abi), developer mode
     5-ml-noble                : CUDA CI, 5-ml-noble, developer mode
 
-    # CUDA wheels (Dockerfile.wheel)
+    # CUDA wheels — x86_64 and SBSA/ARM64 (Dockerfile.wheel)
+    # Architecture is auto-detected via uname -m. The same commands work on
+    # both x86_64 and aarch64 (SBSA) hosts.
     cuda_wheel_py310_dev       : CUDA Python 3.10 wheel, developer mode
     cuda_wheel_py311_dev       : CUDA Python 3.11 wheel, developer mode
     cuda_wheel_py312_dev       : CUDA Python 3.12 wheel, developer mode
@@ -80,9 +82,9 @@ HOST_OPEN3D_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")"/.. >/dev/null 2>&1 && pw
 # Shared variables
 AARCH="$(uname -m)"
 # do cmake pending on the architecture
-CMAKE_VERSION=cmake-3.31.8-linux-${AARCH}
-CUDA_VERSION=12.6.3-cudnn
-CUDA_VERSION_LATEST=12.6.3-cudnn
+CMAKE_VERSION=cmake-3.31.10-linux-${AARCH}
+CUDA_VERSION=12.6.0-cudnn
+CUDA_VERSION_LATEST=12.6.0-cudnn
 
 print_usage_and_exit_docker_build() {
     echo "$__usage_docker_build"
@@ -176,7 +178,11 @@ openblas_build() {
 
 cuda_wheel_build() {
     BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
-    CCACHE_TAR_NAME=open3d-ubuntu-2204-cuda-ci-ccache
+    if [ "${AARCH}" = "aarch64" ]; then
+        CCACHE_TAR_NAME=open3d-ubuntu-2204-cuda-sbsa-ci-ccache
+    else
+        CCACHE_TAR_NAME=open3d-ubuntu-2204-cuda-ci-ccache
+    fi
 
     options="$(echo "$@" | tr ' ' '|')"
     echo "[cuda_wheel_build()] options: ${options}"
@@ -211,6 +217,7 @@ cuda_wheel_build() {
         --build-arg CCACHE_TAR_NAME="${CCACHE_TAR_NAME}" \
         --build-arg CMAKE_VERSION="${CMAKE_VERSION}" \
         --build-arg PYTHON_VERSION="${PYTHON_VERSION}" \
+        --build-arg BUILD_CUDA_MODULE="${BUILD_CUDA_MODULE:?'env var must be set.'}" \
         --build-arg BUILD_TENSORFLOW_OPS="${BUILD_TENSORFLOW_OPS}" \
         --build-arg BUILD_PYTORCH_OPS="${BUILD_PYTORCH_OPS}" \
         --build-arg CI="${CI:-}" \
diff --git a/python/setup.py b/python/setup.py
index 5dbce75297d..d8a283ce119 100644
--- a/python/setup.py
+++ b/python/setup.py
@@ -151,11 +151,11 @@ def finalize_options(self):
 name = "@PYPI_PACKAGE_NAME@"
 with open("README.rst") as readme:
     long_description = readme.read()
-# open3d-cpu wheel for Linux x86_64
+# open3d-cpu wheel for Linux (x86_64 and aarch64/SBSA)
 if "@BUILD_CUDA_MODULE@" == "ON":
     classifiers.append("Environment :: GPU :: NVIDIA CUDA")
 elif (sys.platform.startswith("linux") and
-      platform.machine() in ("i386", "x86_64", "AMD64") and
+      platform.machine() in ("i386", "x86_64", "AMD64", "aarch64") and
       "@BUILD_SYCL_MODULE@" == "OFF"):
     name += "-cpu"
     long_description += ("\n\nThis wheel only contains CPU functionality. "
diff --git a/util/ci_utils.sh b/util/ci_utils.sh
index 092ae3e103a..3b678133df9 100644
--- a/util/ci_utils.sh
+++ b/util/ci_utils.sh
@@ -30,8 +30,9 @@ BUILD_SYCL_MODULE=${BUILD_SYCL_MODULE:-OFF}
 TENSORFLOW_VER="2.20.0"
 TORCH_VER="2.10"
 TORCH_REPO_URL="https://download.pytorch.org/whl/torch/"
+TORCH_NIGHTLY_INDEX_URL="https://download.pytorch.org/whl/nightly"
 # Python
-PIP_VER="25.3"
+PIP_VER="26.0.1"
 PROTOBUF_VER="6.31.1"
 
 OPEN3D_INSTALL_DIR=~/open3d_install
@@ -51,6 +52,7 @@ install_python_dependencies() {
         TF_ARCH_DISABLE_NAME=tensorflow-cpu
         CUDA_VER=$(nvcc --version | grep "release " | cut -c33-37 | sed 's|[^0-9]||g') # e.g.: 117, 118, 121, ...
         TORCH_GLNX="torch==${TORCH_VER}+cu${CUDA_VER}"
+        TORCH_NIGHTLY_INDEX="${TORCH_NIGHTLY_INDEX_URL}/cu${CUDA_VER}"
     else
         # tensorflow-cpu wheels for macOS arm64 are not available
         if [[ "$OSTYPE" == "darwin"* ]]; then
@@ -61,6 +63,7 @@ install_python_dependencies() {
             TF_ARCH_DISABLE_NAME=tensorflow
         fi
         TORCH_GLNX="torch==${TORCH_VER}+cpu"
+        TORCH_NIGHTLY_INDEX="${TORCH_NIGHTLY_INDEX_URL}/cpu"
     fi
 
     python -m pip install -r "${OPEN3D_SOURCE_ROOT}/python/requirements.txt"
@@ -76,10 +79,17 @@ install_python_dependencies() {
     fi
     if [ "$BUILD_PYTORCH_OPS" == "ON" ]; then # ML/requirements-torch.txt
         if [[ "$OSTYPE" == "linux-gnu"* ]]; then
-            python -m pip install -U "${TORCH_GLNX}" -f "$TORCH_REPO_URL"
+            python -m pip install -U "${TORCH_GLNX}" -f "$TORCH_REPO_URL" || {
+                echo "PyTorch ${TORCH_GLNX} not found in stable index, trying nightly..."
+                python -m pip install --pre -U torch --index-url "$TORCH_NIGHTLY_INDEX"
+            }
             python -m pip install tensorboard
         elif [[ "$OSTYPE" == "darwin"* ]]; then
-            python -m pip install -U torch=="$TORCH_VER" -f "$TORCH_REPO_URL" tensorboard
+            python -m pip install -U torch=="$TORCH_VER" -f "$TORCH_REPO_URL" || {
+                echo "PyTorch ${TORCH_VER} not found in stable index, trying nightly..."
+                python -m pip install --pre -U torch --index-url "$TORCH_NIGHTLY_INDEX"
+            }
+            python -m pip install tensorboard
         else
             echo "unknown OS $OSTYPE"
             exit 1
@@ -169,12 +179,16 @@ build_pip_package() {
         echo "Azure Kinect disabled in Python wheel."
         BUILD_AZURE_KINECT=OFF
     fi
-    if [[ "build_jupyter" =~ ^($options)$ ]]; then
+    if [[ "build_jupyter" =~ ^($options)$ ]] && [[ "$AARCH" != "aarch64" ]]; then
         echo "Building Jupyter extension in Python wheel."
         BUILD_JUPYTER_EXTENSION=ON
         BUILD_WEBRTC_FROM_SOURCE=ON
     else
-        echo "Jupyter extension disabled in Python wheel."
+        if [[ "$AARCH" == "aarch64" ]] && [[ "build_jupyter" =~ ^($options)$ ]]; then
+            echo "Jupyter extension not supported on ARM Linux (WebRTC unavailable). Disabling."
+        else
+            echo "Jupyter extension disabled in Python wheel."
+        fi
         BUILD_JUPYTER_EXTENSION=OFF
         BUILD_WEBRTC_FROM_SOURCE=OFF
     fi