Cyclo Intelligence

Open-source full-stack Physical AI platform — data recording, conversion, training, inference, and robot execution in a single repository.

For detailed usage and tutorials, please refer to the documentation below.

• Documentation for AI Worker

Clone

git clone --recurse-submodules https://github.com/ROBOTIS-GIT/cyclo_intelligence.git

Folders at a glance

Folder	Role
shared/	Robot configs, IO helpers, logger
cyclo_brain/	Training + inference (per-backend containers under policy/)
cyclo_data/	Data recording / conversion / hub upload (ROS2 node)
orchestrator/	Session state, UI, behaviour-tree control (ships React UI)
interfaces/	ROS2 msg / srv definitions
docker/	Unified compose, s6-services, Dockerfiles (arm64 / amd64)
docs/	Architecture

Prerequisites

• Docker 24+ with the Compose plugin (docker compose version)
• NVIDIA Container Toolkit configured as the default runtime (docker info | grep "Default Runtime: nvidia")
• ~35 GB free disk for the three pre-built images
• Ports 80, 8100, 9090 free on the host

No Docker Hub login is required — the published images (robotis/cyclo-intelligence, robotis/groot-zenoh, robotis/lerobot-zenoh) are public and pulled anonymously by default.

Quick start (Jetson / ARM64 — same on AMD64)

docker/container.sh start          # pull + start unified image (no rebuild)
docker/container.sh status         # check s6 service state
# UI:          http://localhost/
# control API: http://localhost/api/health
docker/container.sh start-lerobot  # policy on demand (LOAD via UI)
docker/container.sh start-groot    # policy on demand (LOAD via UI)
docker/container.sh stop           # tear everything down

docker/container.sh auto-detects uname -m, so the same commands work on both Jetson and an AMD64 workstation. The default start* flow uses the pre-built image from Docker Hub. Pass --build to rebuild from local Dockerfile (only needed when iterating on Dockerfile changes).

Troubleshooting

• personal access token is expired on pull — A previous docker login left an expired token in ~/.docker/config.json. Run docker logout to clear it; the published images pull anonymously.
• Default Runtime: runc instead of nvidia — Install the NVIDIA Container Toolkit and set nvidia as the default runtime in /etc/docker/daemon.json, then sudo systemctl restart docker.
• /var/run/robotis/agent_sockets/...: permission denied — container.sh will create this with mkdir -p (using sudo if needed). If sudo prompts, grant once or pre-create the directory.

Architecture

See docs/ARCHITECTURE.md for runtime topology and data flow.

Submodules (pinned commit)

• cyclo_brain/sdk/zenoh_ros2_sdk/ ← ROBOTIS-GIT/zenoh_ros2_sdk
• cyclo_brain/policy/lerobot/lerobot/ ← huggingface/lerobot
• cyclo_brain/policy/groot/Isaac-GR00T/ ← NVIDIA/Isaac-GR00T

Related

• AI Worker ROS 2 Packages
• Simulation Models
• Tutorial Videos
• AI Models & Datasets
• Docker Images

License

Apache License 2.0 — see LICENSE.