Pashupati

Autonomous patrol and inspection robot for quadrupeds, legged or wheel-legged, of any brand with a ROS 2 driver: teach and repeat on mapped sites, VLM inspection at stop points, crowd sensing, a handheld controller and the NETRA command center. Built on the Dharma stack.

Packages · Architecture · Quick start · Simulation · Install on the robot · Operate · Develop · Docs

---

Mall corridors are patrolled on fixed routes every shift. Pashupati therefore works by teach and repeat: map a site once, drive a route with its stop points, and the robot repeats it on schedule, inspecting the floor with a vision-language model and counting people along the way. Everything runs on the robot as one supervised service; the Pashupati Controller handheld and NETRA connect to it over the network.

Pashupati runs on Dharma 0.2, the shared autonomy stack: odometry, localization, perception, mapping and sites, the planner, path following, docking, the control mux and health come from the dharma_* repositories, pinned in pashupati.repos. This repository adds what makes it a patrol robot (VLM inspection, people, the NETRA uplink, the handheld gateway) and the bringup with the values tuned on the robot.

Guides: the User Manual for operators and the Tuning Guide for technicians. More in docs.

Packages

Package	Role
pashupati	the robot: profile, stack config, robot.launch.py, sim.launch.py, the end-to-end test
pashupati_inspection	VLM inspection through Ollama (anomaly_detector_node), person detection with YOLO11n ONNX
pashupati_bridge	handheld gateway over UDP (protocol), WebRTC video, RTSP camera source
pashupati_fleet	NETRA uplink: telemetry, stop events, findings, crowd, routes (API), fake_netra
pashupati_utils	what the Pashupati packages share: typed status to protocol names, health figures, launch helpers
dharma_* (pashupati.repos)	utils, interfaces, odometry, localization, perception, mapping, navigation, health, rviz, sim

The robot's driver, URDF and sensors come from a robot package under src/xtras/robots/ (contract); FAST-LIO2 and the Livox driver are submodules in src/xtras/.

Architecture

_{docs/diagrams/system.drawio.svg: open it in draw.io, edit and save, and this picture changes with it.}

• Motion has one way out. control_mux_node is the only publisher of /<robot_ns>/cmd_vel. In AUTO it passes the follower and the planner; in REMOTE the handheld (control/remote/handheld/cmd_vel) or NETRA's dashboard. The protective field, the health speed limit and the latched e-stop apply to every source.
• Patrol is Dharma's path follower on a taught route of the active site. At each stop point it converges, calls the inspection hook (/anomaly_detector_node/check_now), dwells and moves on; a disabled or missing inspection never blocks the patrol.
• Status is typed: the Pashupati nodes read ControlStatus, NavigationStatus, LocalizationStatus, HealthStatus, SiteStatus, StopPointEvent and DockingStatus, and translate them for the handheld protocol v1 and the NETRA API, which keep their wire format.
• Configuration is layered: the robot profile (robot.yaml), the stack as tuned on the robot (stack.yaml), this robot's site file (/etc/pashupati/site.yaml), then launch arguments from /etc/pashupati/pashupati.env.

Hardware

Component	Model	Role
Robot base	any quadruped (legged or wheel-legged) with a ROS 2 driver	robot package, contract
LiDAR	Livox MID-360	FAST-LIO2 odometry, mapping, sectors, safety scan
Camera	Intel RealSense D435i, or an RTSP camera	AprilTags, VLM inspection, people, handheld video
Compute	NVIDIA Jetson Orin NX (JetPack 6)	ROS 2 Humble, Ollama, ONNX Runtime, hardware H.264

Quick start

The dev container has ROS 2 Humble, every dependency and an XFCE desktop over VNC:

git clone --recursive git@github.com:DarmawanRobotics/Pashupati.git && cd Pashupati
scripts/dev.sh up      # build the image, import pashupati.repos into src/dharma, start
scripts/dev.sh build   # colcon build
scripts/dev.sh test    # colcon test
scripts/dev.sh vnc     # TigerVNC localhost:5901 or http://localhost:6080/vnc.html

Without Docker: Ubuntu 22.04 with ROS 2 Humble, then vcs import src < pashupati.repos, rosdep install --from-paths src --ignore-src -y and colcon build.

Simulation

The whole robot runs on dharma_sim, with no robot, Gazebo or GPU:

ros2 launch pashupati sim.launch.py rviz:=true
ros2 run pashupati_fleet fake_netra --port 8765 &       # a local NETRA
ros2 launch pashupati sim.launch.py cc_url:=http://127.0.0.1:8765 inspection:=true
scripts/start_tmux.sh sim                                  # both, plus a status pane

The simulated robot starts on the mall_loop site (in /tmp/pashupati_sim/sites) and localizes itself; teach a route and patrol it with the commands under Operate. The handheld connects with the token sim.

Install on the robot

You need Ubuntu 22.04 with ROS 2 Humble (JetPack 6 on a Jetson), internet access, and the robot's SSH key added to GitHub with read access to the dharma_* repositories.

git clone --recursive git@github.com:DarmawanRobotics/Pashupati.git ~/dev/Pashupati
cd ~/dev/Pashupati
git clone <robot package repo> src/xtras/robots/<name>   # driver + URDF + sensors, src/xtras/README.md
scripts/setup.sh install config          # creates /etc/pashupati/pashupati.env and site.yaml
sudoedit /etc/pashupati/pashupati.env    # ROBOT_ID, ROBOT_DRIVER, ROBOT_NS, SITE, CAMERA, CC_URL
sudoedit /etc/pashupati/site.yaml        # this robot: teleop token, robot.* overrides, tuning
scripts/setup.sh install                 # everything missing: Dharma, SDKs, apt, then the build
scripts/setup.sh install service         # systemd unit, enabled at boot
sudo systemctl start pashupati && journalctl -fu pashupati
scripts/setup.sh                         # every line ok?

Update: git pull && git submodule update --init --recursive && scripts/setup.sh install dharma build config service && sudo systemctl restart pashupati.

Extra	pashupati.env	Install	Gives
Handheld video	WEBRTC=true	scripts/setup.sh install webrtc	H.264 over WebRTC through MediaMTX, encoded only while watched; JPEG without it
VLM inspection	INSPECTION=true, OLLAMA_URL	scripts/setup.sh install ollama	findings at stop points, continuous or on an interval
People	PEOPLE=true	scripts/setup.sh install people_model	YOLO11n ONNX and ONNX Runtime, no PyTorch
RTSP camera	CAMERA=rtsp, CAMERA_URL	—	an IP camera instead of the robot package's camera
NETRA	CC_URL, CC_TOKEN, ROBOT_ID	—	telemetry, findings, routes, remote commands

Pashupati runs on rmw_fastrtps_cpp, domain 0, with ROS_LOCALHOST_ONLY=1, so nothing on the Wi-Fi can bypass the control mux. A shell with the service's environment:

echo "alias pashu='source ~/dev/Pashupati/scripts/env.sh'" >> ~/.bashrc   # once
pashu
pnode pashupati_bridge teleop_udp_node bridge.yaml   # one node, as robot.launch.py runs it

Operate

Teach. Map the site and teach a route (the handheld's Record and Point do the same):

ros2 launch pashupati robot.launch.py mode:=teach site:=mall   # or MODE=teach SITE=mall
ros2 service call /localization/record_tags std_srvs/srv/Trigger           # each tag in view
ros2 service call /route/start std_srvs/srv/Trigger
ros2 service call /route/mark_stop_point dharma_interfaces/srv/MarkStopPoint "{dwell_sec: 10.0}"
ros2 service call /route/stop std_srvs/srv/Trigger    # saved in <site>/routes/, uploaded to NETRA
ros2 service call /mapping/save std_srvs/srv/Trigger

Repeat. Localize (a tag in view: /localization/start), then start from NETRA, from the handheld (Routes → load → Start), or by hand:

ros2 service call /navigation/load_path dharma_interfaces/srv/LoadPath "{path: mall/routes/<date>_<time>.csv}"
ros2 service call /navigation/start std_srvs/srv/Trigger
ros2 service call /navigation/pause std_srvs/srv/Trigger     # resume: /navigation/resume

Drive. REMOTE from the handheld, or ros2 service call /control/set_mode dharma_interfaces/srv/SetMode "{mode: remote, source: handheld}"; REMOTE pauses the patrol, AUTO resumes it.

Emergency stop. Latched in the control mux: the handheld STOP, NETRA, or ros2 service call /control/estop/engage std_srvs/srv/Trigger. Only zero velocity reaches the robot until an explicit /control/estop/release; the patrol then stays paused until resumed.

Check
Health	ros2 topic echo /health/status: level, issues, speed limit, sensor rates, CPU, GPU, RAM, disk
Patrol	ros2 topic echo /navigation/status, /navigation/stop_point
Inspection	ros2 topic echo /anomaly_detector/status: state, model, pace, queue and the last result
People	ros2 topic echo /perception/people_status
Logs	journalctl -fu pashupati; ROS logs older than 7 days are pruned at start

Every Dharma topic and service: the hub's Topics and Services.

Develop

Pashupati/
├── .github/workflows/      ci.yml (lint, scripts, docs, whole stack on Humble), people-model.yml
├── .devcontainer/ compose.yaml docker/   dev container (Humble, VNC)
├── docs/                   User Manual, Tuning Guide (Markdown, PDFs built from them), diagrams
├── pashupati.repos         the Dharma versions (vcstool)
├── scripts/                setup.sh, run_robot.sh (service), env.sh, dev.sh, start_tmux.sh
└── src/
    ├── pashupati pashupati_*   the product packages
    ├── dharma/                 imported from pashupati.repos (git-ignored)
    └── xtras/                  submodules: livox_ros_driver2, FAST_LIO; robots/<robot>, slam/leg_kilo

colcon test runs the style check and the unit and in-process ROS tests of every package, and src/pashupati/test/e2e the whole robot on the simulator: it localizes, teaches a route with three stop points, patrols it with a stand-in VLM answering every inspection, hands over to the handheld and back through the mux, and reports to a local NETRA. CI runs all of it on every push and pull request. Branch, commit and code rules are in CONTRIBUTING.md, changes in CHANGELOG.md.

Pashupati is released under the Apache License 2.0.