Cyclo — ROBOTIS Open-Source Modular Physical AI Framework

TL;DR

Cyclo is a modular Apache-2.0 framework from ROBOTIS (the Korean maker of the DYNAMIXEL actuator series and open ROS platforms like TurtleBot and OpenMANIPULATOR) for building end-to-end Physical AI systems on their hardware lineup. It bundles four public sub-repositories under one umbrella — cyclo_manager (operations), cyclo_intelligence (imitation-learning workflows), cyclo_control (whole-body control), and cyclo_lab (simulation + RL) — plus shared interfaces (robotis_interfaces) and third-party integrations (robotis_applications). The stated goal is to put the building blocks of Physical AI in the open — data collection through deployment — so integrators aren’t locked into a single-vendor closed path, with an optional private-extension stack (cyclo_supervisor, cyclo_hub) plugging alongside for proprietary supervision/data.

Key claims

Cyclo is organized as four public modules covering the full Physical AI pipeline — Manager (operations/system management), Intelligence (next-generation imitation-learning workflows), Control (whole-body control and execution), and Lab (simulation + reinforcement learning) — each in a separately versioned GitHub repository under the ROBOTIS-GIT org [§Modules & repositories].
Two additional shared repositories back the modules: robotis_interfaces for cross-module message/type definitions, and robotis_applications for application-level integrations and third-party dependencies [§Modules & repositories].
The framework targets the ROBOTIS Physical AI product line — AI Sapiens, AI Worker, OpenMANIPULATOR, and ROBOTIS Hand — and the DYNAMIXEL actuator series (DXL, DXL-X, DXL-P, DXL-Y, DXL-Q) via the existing DynamixelSDK interface stack [§Physical AI Lineup, §Actuator products].
The design explicitly reserves a private extension surface — Cyclo Supervisor (robot supervision) and Cyclo Hub (private data infrastructure) — that plugs into the public stack rather than being open-sourced [§Private stack].
Licensing is Apache 2.0 across the public modules [§License badge].

Method

Cyclo is a coordinating umbrella repository — it contains a README, a framework diagram (assets/cyclo_framework.png), and pointers to the four functional modules; the actual code lives in the sub-repositories. Physical AI is decomposed into five vertical concerns (AI integration, action-data workflows, robot control, simulation, and operations), each owned by one Cyclo module. cyclo_lab covers the sim-and-RL layer; cyclo_intelligence covers imitation-learning workflows (data → training); cyclo_control executes on real hardware via the DynamixelSDK and product-specific stacks (AI Sapiens, AI Worker, OpenMANIPULATOR, ROBOTIS Hand); cyclo_manager handles system-level orchestration. Contributions are routed to the individual sub-repositories.

Results

No benchmarks are reported — this is an infrastructure release, not a model or policy release. The claim is architectural: a single vendor’s actuators + robot platforms + software stack are being opened as a coordinated Physical AI framework rather than as isolated per-product repositories, with a stated private-extension slot for proprietary supervision and data.

Why it’s interesting

Sits in the “infrastructure-tier open release” pattern already tracked on Open foundation-model releases alongside Genesis World 1.0 (Genesis World — Simulation platform for general-purpose robotics & embodied AI learning, The Role of Simulation in Scalable Robotics, Genesis World 1.0, and the Path Forward) — both are Apache-2.0 releases where the artifact is a coordinated stack (physics + control + tooling) rather than weights, but where Genesis stakes out the sim/renderer/compiler position, Cyclo stakes out the vendor-of-record hardware position, wrapping ROBOTIS’s existing product line (DYNAMIXEL, OpenMANIPULATOR, AI Worker) into one framework. Complements the vertically-integrated embodied stacks from Chinese labs — Alibaba’s Qwen-Robot Suite (Qwen-Robot Suite launch — Qwen-RobotManip, Qwen-RobotNav, Qwen-RobotWorld + RobotClaw framework (Alibaba Qwen)) and RynnBrain 1.1 (RynnBrain 1.1 — Open Embodied Foundation Models (Alibaba DAMO Academy)), Tencent’s Hy-Embodied line (RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination, Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack) — but from a hardware-first vendor rather than a model-lab vendor, and with an explicit private-supervision extension slot rather than fully open. The private/public split is worth watching: it’s a concrete instantiation of the release-pattern question the Open foundation-model releases concept has been tracking (what’s the “minimum viable public surface” that still lets integrators build without needing the vendor’s internal tools).