ROS 2 Jazzy · ros2_control · bimanual hardware

Dual-Arm UR3 / UR3e + Allegro Hand Drivers

The workspace that brings up a bimanual torso: two Universal Robots arms, two 16-DoF Allegro Hand V4 over PCAN, three RealSense cameras, a calibrated URDF with collision geometry, and bridge nodes so any program (a learned policy, a grasp planner, teleoperation) can drive the cell with joint-position topics.

Code & setup guide Runs on it: dexterous pick-and-place Runs on it: 6D object tracking

Left to right: the real cell, its URDF twin from the same calibrated camera pose, and the overlay (red = real table edge, cyan = twin).

2 armsUR3 (right) + UR3e (left) in one ros2_control controller manager

2 × 16 DoFAllegro Hand V4, customised driver over PCAN-USB

44 jointsmerged into one /joint_states at 100 Hz

3 camerasRealSense D415 with calibrated intrinsics and extrinsics

What one launch file brings up

armsUR driver + ros2_controlBoth arms' hardware interfaces, forward-position / trajectory / velocity controllers, dashboard and safety stopper per arm.

handsAllegro V4 driverPID joint-position control over CAN, with a PWM ceiling that keeps the thumb motor from overheating.

camerasRealSense, patchedColour + depth per camera. Calibrated intrinsics load from YAML via camera_info_url, and extrinsics are published as static TFs to world.

bridgesTopic-level controlMerged joint states, arm streaming with damping, and hand streaming that releases when its input topic goes silent.

Robot description: visual and collision

The URDF is generated by xacro at launch: world, T-shaped stand and table, a UR3e and a UR3 with their factory kinematic calibrations, and an Allegro hand on each flange (prefixes la_ / ra_). The same model is exported to Isaac Lab for sim-to-real work.

Collision model — Collision geometry: right arm blue, left arm orange, stand and table grey

Collision model front — Front view, collision

Rendered digital twin — The same cell rendered in Isaac Sim

Right camera overlay — Second camera: real, twin, overlay

On the real robot

Pick-and-place over 10 objects, driven through these drivers (right arm + right Allegro), with FoundationPose perception and BODex grasp synthesis.

From dexterous-pick-and-place (94/100 hardware trials), 3× speed.

Hardware

component	model	connection	prefix
right arm	UR3	Ethernet 192.168.0.120, External Control	`ra_`
left arm	UR3e	Ethernet 192.168.0.121, External Control	`la_`
right hand	Allegro Hand V4 (P231)	PCAN-USB, channel 33	`ra_`
left hand	Allegro Hand V4 (P249)	PCAN-USB, channel 32	`la_`
cameras	RealSense D415 ×3	USB 3	`front_right`, `extra`, `front_left`

Try it

git clone https://github.com/ismael-esp/dual-arm-allegro-ros2.git && cd dual-arm-allegro-ros2
bash scripts/install_pcan_host.sh      # host: PEAK pcan kernel module
bash docker/build.sh                   # ROS 2 Jazzy + ros2_control + UR + MoveIt + librealsense + PCAN libs
bash docker/run.sh                     # repo mounted at /home/dev/ros2_ws
bash scripts/setup_workspace.sh        # pinned upstream drivers + RealSense patch + colcon build
source install/setup.bash
ros2 launch my_dual_robot_cell_control start_system_drivers.launch.py      # real robot
ros2 launch my_dual_robot_cell_control start_robots.launch.py \
    la_use_mock_hardware:=true ra_use_mock_hardware:=true                  # no hardware

The README covers robot-side setup, commanding the cell, configuration, calibration and troubleshooting.