Skip to content

INSTALL.md

Latest commit

 

History

History
144 lines (105 loc) · 5.44 KB

INSTALL.md

File metadata and controls

144 lines (105 loc) · 5.44 KB

Installation

Prerequisites

Requirement Tested with
Ubuntu 24.04
ROS 2 Jazzy
NVIDIA GPU with sufficient VRAM (24GB+) + CUDA CUDA 12.x
Python 3.10+ (pyenv or system)

ROS2

for installing ROS2 refer to https://docs.ros.org/en/jazzy/Installation.html .

PyTorch

Make sure to match the cu suffix in ./requirements.txt to your CUDA version (tested with cu128).

Workspace Setup

source /opt/ros/${ROS_DISTRO}/setup.bash
mkdir -p ~/ros2_ws/src && cd ~/ros2_ws/src
git clone git@github.com:MIT-SPARK/DAAAM.git daaam
bash daaam/install/install.sh

The script clones all 17 repos, installs system & rosdep deps, writes colcon_defaults.yaml, builds the C++ workspace, and runs pip install -r requirements.txt && pip install -e . for the Python package.

Manual Step-by-Step

source /opt/ros/${ROS_DISTRO}/setup.bash

# System deps
sudo apt install python3-vcstool python3-tk libgoogle-glog-dev \
  nlohmann-json3-dev glpk-utils libglpk-dev ros-dev-tools

# Clone workspace repos
cd ~/ros2_ws/src
vcs import . < daaam/install/packages.yaml --workers 1 --skip-existing

# Rosdep
rosdep install --from-paths . --ignore-src -r -y

# Colcon defaults (write once)
cat > ~/ros2_ws/colcon_defaults.yaml <<'YAML'
---
build:
  symlink-install: true
  packages-skip: [khronos_msgs, khronos_ros, khronos_eval, hydra_multi_ros, spark_fast_lio, ouroboros_ros, ouroboros_msgs]
  cmake-args:
    - --no-warn-unused-cli
    - -DCMAKE_BUILD_TYPE=RelWithDebInfo
    - -DCONFIG_UTILS_ENABLE_ROS=OFF
    - -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
    - -DGTSAM_USE_SYSTEM_EIGEN=ON
YAML

# Build (gtsam is RAM-hungry — use -j2 on <16 GB machines)
cd ~/ros2_ws
colcon build --continue-on-error

# Python deps + editable install
cd ~/ros2_ws/src/daaam
pip install -r requirements.txt
pip install -e .

Workspace Repo Map

All repos live side-by-side under ~/ros2_ws/src/. Apart from daaam and daaam_ros these are all hydra dependencies. They will be installed automatically if you follow the instructions above.

Path Repo Branch Notes
daaam MIT-SPARK/DAAAM main Core library (this repo)
daaam_ros MIT-SPARK/DAAAM-ROS main ROS 2 interface
hydra MIT-SPARK/Hydra project/daaam 3D Dynamic Scene Graphs
hydra_ros MIT-SPARK/Hydra-ROS project/daaam Hydra ROS 2 wrapper
spark_dsg MIT-SPARK/Spark-DSG project/daaam Scene graph data structure
khronos MIT-SPARK/Khronos main Spatio-temporal reasoning
config_utilities MIT-SPARK/config_utilities main SPARK config helpers
ianvs MIT-SPARK/Ianvs main SPARK utilities
kimera_pgmo MIT-SPARK/Kimera-PGMO ros2 Mesh optimization
kimera_rpgo MIT-SPARK/Kimera-RPGO develop Robust PGO
pose_graph_tools MIT-SPARK/pose_graph_tools ros2 Pose graph msgs
semantic_inference MIT-SPARK/semantic_inference main Semantic segmentation
spatial_hash MIT-SPARK/Spatial-Hash main Spatial indexing
teaser_plusplus MIT-SPARK/TEASER-plusplus master Registration
gtsam borglab/gtsam release/4.2 Factor graphs
small_gicp koide3/small_gicp master Fast ICP
vision_opencv ros-perception/vision_opencv rolling cv_bridge for ROS 2

Optional: TensorRT Acceleration

FastSAM and the BotSort ReID model can be exported to TensorRT .engine files for real-time inference. Engine files are GPU-specific and TensorRT-version-specific — they must be rebuilt when switching GPUs or updating TensorRT.

Prerequisites

The PyTorch CUDA version and the TensorRT CUDA version must match. If PyTorch is installed with cu128, TensorRT must the same version of CUDA that your GPU is running.

Warning: If you are using a CUDA version other than 12.X or do not intend to use TRT acceleration, adjust the version of tensorrt-cuXXin requirements.txt .

Further, the defaults in all launch files in DAAAM-ROS are set to .engine files. if you intend to use standard .pt models, adapt the launch files.

Exporting FastSAM TRT model

python scripts/export_fastsam_trt.py --model_name FastSAM-x

Exporting BotSort ReID (vanilla CLIP) TRT model

python scripts/export_vanilla_clip_engine.py

Configuration

Set the engine paths in the launch file and/or config/pipeline_config.yaml:

segmentation:
  model_name: "fastsam/FastSAM-x-640x480.engine"
  imgsz: [480, 640]

tracking:
  reid_weights: "checkpoints/reid_weights/clip_general.engine"

When using .pt files instead (no TensorRT), the code auto-detects the backend from the file extension — no other changes needed.

Optional: GLPK for Assignment Optimization

The min_frames_max_size assignment worker uses CVXPY to solve a mixed-integer program. It requires the GLPK solver:

# Ubuntu/Debian
sudo apt install glpk-utils libglpk-dev
pip install cvxopt   # provides cp.GLPK_MI to CVXPY

If GLPK_MI is unavailable at runtime, the worker falls back to cp.SCIPY (scipy.optimize.milp). Note that scipy >= 1.15.0 ships HiGHS 1.8 which regresses significantly on this problem structure — solve times 5-7x slower than scipy 1.14.x. Pin scipy<1.15 if GLPK is not an option.