Skip to main content

AR4 Physical-AI

Repo: aegean-ai/ar4-physical-ai (private) Environments: ros.dev.gpu, torch.dev.gpu Status: Epic in progress Beads: auraison-rj7 (epic), auraison-azf, auraison-dgd, auraison-b7v, auraison-4b0, auraison-dkb Jira: AURA-342

Overview​

6-DOF robotic arm (AR4 MK3) platform for Vision-Language-Action (VLA) manipulation. Uses LeRobot for recording, training, and inference; GR00T N1 as a future policy backend; Gazebo Harmonic for simulation; MoveIt2 + ros2_control for safe trajectory execution.

Architecture​

  • ros.dev.gpu: ROS 2 Jazzy + Gazebo Harmonic simulation, MoveIt2 motion planning, gz_ros2_control
  • torch.dev.gpu: VLA inference via vLLM + Zenoh queryable, policy training
  • Zenoh: middleware for non-ROS transport between inference server and ROS control loop

Critical path​

azf (URDF + Gazebo sim)
  +---> dgd (MoveIt2 config)  ---> dkb (skill library)
  |
  +---> b7v (Policy Server)   ---> 4b0 (episode logging) ---> drp (GR00T N1)

Datasets​

  • Episode recordings stored in data-plane lakehouse (LeRobot format)
  • Synthetic data via GR00T-Dreams + Cosmos pipeline (v2)

Platform services​

  • User plane: Ray Jobs for training, inference server as long-running Ray actor
  • Data plane: Episode datasets, model checkpoints in warehouse/
  • Control plane: NotebookAgent for training job submission