These are the simulation models of the robots in the Cyber-physical Systems Lab at Uppsala University. All of them live in this one repository. The meshes, URDFs, USDs and MJCFs here are the only copies to use. The same URDF works in RViz2, Gazebo Fortress, MuJoCo and Isaac Sim / Isaac Lab. Each simulator can run from its own conda environment or its own Docker image.
Robots_IsaacLabhas been merged into this repo and is archived. See docs/MIGRATION.md.
Robot (robots.yaml key) |
Description | RViz2 | Gazebo Fortress | MuJoCo | Isaac Sim / Lab |
|---|---|---|---|---|---|
go2 |
Unitree Go2 quadruped | ✅ | 🟢 | 🟢 ¹ | 🟢 ² |
myagv |
Elephant myAGV base | ✅ | 🟢 | 🟢 | 🟢 ³ |
mycobot_280_pi |
myCobot 280 Pi + camera flange + pump (+3 variants) | ✅ | 🟢 | 🟢 | 🟢 |
mycobot_280_gripper |
myCobot 280 + adaptive gripper | 🟢 | 🟢 | 🟢 | 🟢 |
mycomb |
myAGV + myCobot 280 Pi + gripper (mycomb1 = second instance) |
✅ | 🟢 | 🟢 | 🟢 |
nova5 |
Dobot Nova 5 | 🟢 | 🟢 | 🟢 | 🟢 |
nova5_2f85 |
Dobot Nova 5 + Robotiq 2F-85 | ✅ | 🟢 | 🟢 | 🟢 |
✅ verified by a lab member · 🟢 loads, holds its pose and renders correctly in an automated run (see the gallery; captured 2026-11-05 with RViz2 Humble, Gazebo Fortress 6.18, MuJoCo 3.x, Isaac Sim 6.0) · 🟡 loads, but uses legacy assets · ⬜ not yet verified · ❌ known broken
Physics everywhere still uses placeholder masses for the Elephant robots (see Known issues).
¹ For serious Go2 work, use the tuned MuJoCo Menagerie Go2.
² Isaac Lab ships UNITREE_GO2_CFG, and isaaclab/cps_robots_cfg.py uses it.
³ myAGV has no movable joints, so Isaac imports it as a rigid prop, not an articulation.
CI (.github/workflows/validate.yml) runs on every push:
- the static URDF checks
check_urdf(urdfdom)- Gazebo URDF→SDF
- a MuJoCo load-and-step test
To turn a cell ✅, a person must still open the robot in that simulator. See Validation.
Every robot is shown in every simulator from the same camera angle (from the +x+y diagonal, 25° above). The images live in docs/screenshots/<simulator>/<robot>.png. Click one to enlarge it.
| Robot | RViz2 | Gazebo Fortress | MuJoCo | Isaac Sim / Lab |
|---|---|---|---|---|
go2 |
![]() |
![]() |
![]() |
![]() |
myagv |
![]() |
![]() |
![]() |
![]() |
mycobot_280_pi |
![]() |
![]() |
![]() |
![]() |
mycobot_280_gripper |
![]() |
![]() |
![]() |
![]() |
mycomb |
![]() |
![]() |
![]() |
![]() |
nova5 |
![]() |
![]() |
![]() |
![]() |
nova5_2f85 |
![]() |
![]() |
![]() |
![]() |
28/28 screenshots captured. Regenerate: see Screenshots.
All_Robot/
├── robots.yaml # registry: the one place that lists every robot
├── robots/<name>/
│ ├── urdf/<name>.urdf # canonical URDF, package://cps_robot_description/... mesh URIs
│ ├── urdf/standalone/ # generated: relative mesh paths (Isaac Sim, MuJoCo, any non-ROS tool)
│ ├── urdf/isaac/ # generated: standalone + flattened meshes, input of isaaclab/convert_all.sh
│ ├── meshes/ # DAE / STL (+ meshes/flat/: generated STL of each DAE, see below)
│ ├── mjcf/<name>.xml, scene.xml # generated MuJoCo model
│ └── usd/<name>.usd # generated Isaac USD (usd/legacy/ = old hand-made USDs)
├── envs/ # conda envs: env_isaaclab, env_mujoco, env_ros_gz + setup.sh
├── docker/ # one image per simulator + compose.yaml + run.sh
├── launch/ # ROS 2: display, all_in_one (RViz2), gazebo (Fortress)
├── isaaclab/ # Isaac Lab ArticulationCfg, URDF->USD conversion, spawn/check script
├── scripts/ # normalize_urdf, build_assets, validate, migrate, screenshot_*, make_gallery
├── package.xml, CMakeLists.txt # repo root is the ROS 2 package `cps_robot_description`
└── docs/screenshots/<sim>/ # rviz, gazebo, mujoco, isaac -> README gallery
Rule: edit only the canonical urdf/<name>.urdf and meshes/. Everything else is generated by scripts/build_assets.py or isaaclab/convert_all.sh.
git clone https://github.com/Cyber-physical-Systems-Lab/All_Robot.git && cd All_Robot| Simulator | Conda environment | Docker image | Needs |
|---|---|---|---|
| Isaac Sim 4.5 + Isaac Lab 2.1 | env_isaaclab |
cps/isaaclab:4.5.0-2.1.0 |
Ubuntu 22.04, NVIDIA RTX GPU, driver ≥ 535 |
| MuJoCo 3.x | env_mujoco |
cps/mujoco:3 |
any CPU |
| ROS 2 Humble + Gazebo Fortress + RViz2 | env_ros_gz (RoboStack) |
cps/ros_gz:humble-fortress |
any CPU |
Conda setup. You need Miniconda; mamba is used automatically when it's installed. Each command can be re-run safely.
bash envs/setup.sh mujoco # -> env_mujoco
bash envs/setup.sh ros_gz # -> env_ros_gz (also colcon-builds this package into ~/cps_ws)
bash envs/setup.sh isaaclab # -> env_isaaclab (clones Isaac Lab v2.1.0 to ~/IsaacLab, ~25 GB)Docker setup. You need Docker with Compose v2. Isaac also needs the NVIDIA Container Toolkit.
docker/run.sh build mujoco ros_gz isaaclab # once; isaaclab pulls nvcr.io/nvidia/isaac-sim:4.5.0 (~20 GB)
docker/run.sh <mujoco|ros_gz|isaaclab> [cmd] # repo is mounted, X11 + GPU forwarded# conda
conda activate env_ros_gz
ros2 launch cps_robot_description display.launch.py robot:=nova5_2f85
ros2 launch cps_robot_description display.launch.py robot:=mycobot_280_pi variant:=with_pump
ros2 launch cps_robot_description all_in_one.launch.py # all robots side by side
# docker
docker/run.sh ros_gz ros2 launch cps_robot_description display.launch.py robot:=go2With a system ROS install instead, put the repo in a colcon workspace and run colcon build --packages-select cps_robot_description.
# conda
conda activate env_ros_gz
ros2 launch cps_robot_description gazebo.launch.py robot:=nova5_2f85
ros2 launch cps_robot_description gazebo.launch.py robot:=go2
# docker
docker/run.sh ros_gz ros2 launch cps_robot_description gazebo.launch.py robot:=nova5_2f85- Arms (
fixed_base: true) are welded to the world. Mobile robots spawn free atspawn_z. - The launch file sets
IGN_GAZEBO_RESOURCE_PATHso thatpackage://cps_robot_description/...resolves. - Every actuated joint gets a Gazebo
JointPositionControllerthat holds the start pose (init_jointsinrobots.yaml, else 0), so arms stay up and Go2 stands. Mimic joints are left free. - Command a joint from ROS:
ros2 topic pub --once /nova5/joint3/cmd_pos std_msgs/msg/Float64 "{data: 1.0}"(topic/<robot>/<joint>/cmd_pos). Joint states come back on/joint_states, sorobot_state_publisherfollows the simulation. - Collision meshes are swapped for the flattened STLs in
meshes/flat/: Gazebo's DART/ODE loader crashes on several of the Collada files (and on any OBJ). Visuals stay DAE.
The MJCF files are generated from the canonical URDFs. DAE meshes are flattened to binary STL in robots/<r>/meshes/flat/, because MuJoCo can't read DAE (units and node transforms baked in, <up_axis> ignored like RViz does).
# conda
conda activate env_mujoco
python scripts/build_assets.py --mujoco # generate (all robots, or list names)
python -m mujoco.viewer --mjcf=robots/nova5_2f85/mjcf/scene.xml
# docker
docker/run.sh mujoco python scripts/build_assets.py --mujoco
docker/run.sh mujoco python -m mujoco.viewer --mjcf=robots/nova5_2f85/mjcf/scene.xmlWhat the generator adds:
- A
freejointfor mobile robots. equalityconstraints for URDF<mimic>joints (grippers).- A
positionactuator on every joint (mimic joints excluded). scene.xml, with a floor and a light.
Use robots/<r>/mjcf/<r>.xml with <include> in your own scenes.
Tested with Isaac Sim 4.5 + Isaac Lab 2.1 (the Docker image) and Isaac Sim 6.0 + current Isaac Lab (the 2026-11-05 screenshots). The generated USDs are committed, so you can skip convert_all.sh unless you change a URDF.
# conda (envs/setup.sh set ISAACLAB_PATH and CPS_ROBOTS_PATH inside the env)
conda activate env_isaaclab
bash isaaclab/convert_all.sh # robots/<r>/urdf/isaac/*.urdf -> robots/<r>/usd/<r>.usd
python isaaclab/spawn_robot.py --robot nova5_2f85 # view + wiggle joints
bash scripts/screenshot_isaac.sh # PASS/FAIL table + screenshots, one process per robot
# docker (inside the container, Isaac Lab is at /workspace/isaaclab)
docker/run.sh isaaclab bash isaaclab/convert_all.sh
docker/run.sh isaaclab /workspace/isaaclab/isaaclab.sh -p isaaclab/spawn_robot.py --robot nova5_2f85On its first start, Isaac Sim pulls extensions and compiles shaders, which takes about 10 minutes. With Docker, these are cached in named volumes.
Notes on the conversion:
convert_all.shreadsurdf/isaac/, made bypython scripts/build_assets.py --isaac. It is the standalone URDF with every Collada mesh that uses<instance_node>(the Onshape-exported myCobot / myAGV parts) replaced by its flattened STL. Isaac's URDF importer drops those node transforms and would pile the parts onto the link origin.- Newer Isaac Lab converters write a folder (
usd/<r>.usd/<urdf-stem>/<urdf-stem>.usda+payloads/) instead of one file.cps_robots_cfg.pyaccepts both. - Mimic joints (grippers) are never driven: the importer turns them into PhysX mimic constraints, and an actuator on them as well makes the solver produce NaNs.
- Run several robots in separate processes (
scripts/screenshot_isaac.shdoes). In one process they leak state into each other.
Isaac Sim GUI only: go to File → Import and pick robots/<r>/urdf/standalone/<r>.urdf. Importer options:
- Fix Base Link: on for arms.
- Joint Drive: Position.
- Merge Fixed Joints: off.
In your own Isaac Lab project:
import os, sys; sys.path.append(os.path.join(os.environ["CPS_ROBOTS_PATH"], "isaaclab"))
from cps_robots_cfg import make_cfg
robot_cfg = make_cfg("nova5_2f85").replace(prim_path="{ENV_REGEX_NS}/Robot")make_cfg resolves the USD in this order:
usd/<r>.usd(regenerated).usd/legacy/...(the old hand-made USD).- For
go2, Isaac Lab's built-inUNITREE_GO2_CFG.
The actuator gains are in ACTUATORS / GAINS in isaaclab/cps_robots_cfg.py. The start pose comes from init_joints in robots.yaml.
python scripts/validate.py -v # static: tree, meshes exist, inertia, limits, no /home paths
python scripts/validate.py --mujoco # in env_mujoco / docker mujoco
python scripts/validate.py --gazebo # in env_ros_gz / docker ros_gz (runs `ign sdf -p`)
bash scripts/screenshot_isaac.sh # in env_isaaclab / docker isaaclabA cell in the status table becomes ✅ only after a lab member has:
- Run the matching check.
- Opened the robot in the simulator and seen that the meshes are in the right place and the joints move the right way.
Write the simulator version and conda/Docker in the PR description.
The gallery above is generated, so don't paste images into the README by hand.
bash scripts/capture_all.sh # all four simulators via Docker, then refresh the gallery
bash scripts/capture_all.sh mujoco isaac # a subsetOr, inside each conda environment:
| Simulator | Command (from the repo root) | How it renders |
|---|---|---|
| RViz2 | bash scripts/screenshot_ros.sh rviz (env_ros_gz) |
Xvfb virtual display |
| Gazebo | bash scripts/screenshot_ros.sh gazebo (env_ros_gz) |
Xvfb + software OpenGL, GUI reduced to the 3D view |
| MuJoCo | MUJOCO_GL=egl python scripts/screenshot_mujoco.py (env_mujoco) |
offscreen renderer |
| Isaac Sim / Lab | bash scripts/screenshot_isaac.sh (env_isaaclab) |
RTX camera sensor |
The ROS script needs the package built and sourced (colcon build in a workspace containing this repo). Then run python3 scripts/make_gallery.py and commit docs/screenshots/ together with README.md.
All four use 960×720 images, a 45° vertical field of view and the camera from robots.yaml, so the pictures line up.
From CI: every push uploads CPU-rendered RViz2, Gazebo and MuJoCo screenshots as a workflow artifact. To commit them instead (for example after changing a robot on a machine without the simulators), go to GitHub → Actions → validate-models → Run workflow and tick commit_screenshots. Isaac needs an RTX GPU, so capture it locally.
To change the framing of a robot, edit camera: in robots.yaml. All four simulators use it.
- Put the files in
robots/<name>/meshes/androbots/<name>/urdf/<name>.urdf. - Add an entry to
robots.yaml(copy an existing one). python scripts/normalize_urdf.py <name>: fixes mesh URIs and fills missing inertia and limits. Every change is listed.python scripts/build_assets.py --mujoco --isaac <name>, thenbash isaaclab/convert_all.sh <name>.python scripts/validate.py -v <name>. Add a row to the status table above. CI checks the rest.- Add a
camera:entry inrobots.yamland capture screenshots (see Screenshots).
-
Placeholder physics. These links had no
<inertial>, so they got a placeholder mass fromnormalize_urdf.py, marked with an XML comment:- myAGV
- myCobot 280 (all variants)
- the
mycombarm
Several joints also had
effort="0" velocity="0"and were given the defaults fromrobots.yaml. Seedocs/normalize_report.md. Replace these with CAD or datasheet values before using the models for dynamics or RL. -
myAGV is a visual shell with no wheels or drive.
-
DAE → STL flattening applies the Collada
<unit>scale and node transforms, and ignores<up_axis>like RViz. The myCobot M5 meshes were taggedY_UPalthough their coordinates are Z-up; they are now taggedZ_UP, so Isaac and RViz agree. -
Large meshes (>10 MB DAE) slow down Gazebo and MuJoCo collision. Simplified collision meshes would help.
-
RTX 50-series GPUs need a PyTorch nightly build in
env_isaaclab; see the note inenvs/env_isaaclab.yml.
The meshes and URDFs come from, or are adapted from, the vendor and community sources below. Third-party assets keep their original licenses.
| Robot | Source |
|---|---|
| Go2 | Unitree Robotics – unitree_ros/go2_description |
| myAGV | Elephant Robotics, URDF adapted in-lab |
| myCobot 280 Pi / gripper | elephantrobotics/mycobot_ros, mycobot_ros2, extended with camera flange and suction pump |
| Nova 5 | Dobot-Arm/DOBOT_6Axis_ROS2_V3 |
| Robotiq 2F-85 | ros-industrial/robotiq |
Composites (mycomb, nova5_2f85) |
Composed by Xuezhi Niu |
If any asset is missing attribution, please open an issue.
MIT for this repository's own code and composed models (see LICENSE). Vendor meshes keep their original licenses.





























