A π0 vision-language-action grasping policy fine-tuned on 200k teleop episodes, served as a ROS 2 action with a runtime safety envelope (force, workspace, timeout). Bring your own parallel or suction gripper.
π0 Grasping Skill (VLA)
Pre-trained π0 grasping policy with ROS 2 skill server and safety envelope.
A π0 vision-language-action grasping policy fine-tuned on 200k teleop episodes, served as a ROS 2 action with a runtime safety envelope (force, workspace, timeout). Bring your own parallel or suction gripper.
- Cluttered bin grasping
- Goods-to-person picking
- Lab automation
Serves the π0 grasping policy as a ROS 2 action server with workspace + force safety envelope.
- Workstation or on-robot box with ≥ RTX 4070 / Jetson Thor
- ROS 2 Humble + MoveIt 2 servo running for the target arm
- Wrist-mounted RGB camera publishing /wrist_cam/color
- RosLab model licence — accept EULA on first run
- 01Install the skill server
pip install roslab-pi0-skill==0.6.2 ros2 run roslab_pi0 accept_eula - 02Download the policy weights (≈3.4 GB)
roslab-pi0 fetch --variant grasp-v2 --device cuda - 03Start the skill server with a safety envelope
ros2 launch roslab_pi0 grasp_server.launch.py \ arm:=fr3 gripper:=robotiq_2f_85 \ workspace_xyz:="[-0.4,0.4,-0.3,0.3,0.0,0.6]" max_force_n:=20 - 04Issue a grasp goal
ros2 action send_goal /pi0_grasp roslab_pi0/action/Grasp \ "{prompt: 'pick the red mug', timeout_s: 8.0}"
Inference latency under 35 ms on RTX 4070; success rate logged to /pi0_grasp/metrics.
ros2 topic echo /pi0_grasp/metrics --onceA hardened ROS 2 Nav2 distribution with behavior-tree recoveries, multi-floor map switching, costmap plugins for dynamic obstacles, and Foxglove-ready telemetry. One-command deploy to any ROS 2 Humble or Jazzy robot.
Production-tuned ROS 2 Nav2 stack with BT-based recovery and multi-floor maps.
A library of MoveIt 2 task plans for bin picking, kitting and machine tending. Ships with calibration routines for Franka FR3, UR-series and legacy OEM arms, plus a skill-runner that exposes each routine as a ROS 2 action.
Drop-in MoveIt 2 skills: bin picking, kitting, machine tending — calibrated per arm.
A pre-integrated Isaac ROS perception graph — ESS depth, segmentation, AprilTag pose, and pedestrian detection — packaged as a single Docker image with sample launch files for Jetson AGX Orin and IGX.
NVIDIA Isaac ROS perception bundle: depth, segmentation, AprilTag, person detection.