Robonix — The Embodied AI Operating System

Rust-native EAIOS. Skill-centric runtime with Liaison, Pilot, Atlas, Nexus, Executor — human intent to embodied action.

简体中文 · Docs (mdBook) · Rust workspace · Examples & E2E

Quickstart

git clone --recursive https://github.com/syswonder/robonix
cd robonix/rust
make install                              # → ~/.cargo/bin (rbnx, robonix-atlas, …)
pip install -r examples/requirements.txt
cp examples/.env.example examples/.env    # fill VLM_API_KEY + VLM_BASE_URL + VLM_MODEL
./examples/run.sh                         # atlas + vlm + tiago sim + pilot + …

Open another terminal and talk to the agent:

rbnx chat

Full first-run walkthrough: docs/src/getting-started/quickstart.md.

What's here

rust/	Cargo workspace — atlas / pilot / executor / liaison / cli / codegen / sdk / interfaces / buffer
rust/contracts/	Stable contract_id definitions (TOML). Map to ROS IDL under rust/crates/robonix-interfaces/lib/.
rust/examples/packages/	Example packages: vlm_service, memsearch_service, tiago_sim_stack, maniskill_vla_demo, zero_copy_demo, clawhub_skills
docs/	mdBook — architecture, interface catalog, integration guides
images/	Logo + demo assets

Architecture — 30 seconds

Four-layer EAIOS abstraction: primitive → service → skill → task. A unified control plane (Atlas) handles registration, discovery, and channel negotiation across transports (gRPC, MCP, ROS 2, shared memory). Agent reasoning lives in Pilot (VLM + ReAct + TaskGraph), tool dispatch in Executor, user-facing interaction in Liaison.

Dive deeper:

• Overview — control/data plane, one full request end-to-end
• Crates — each binary's role and port
• Namespaces & contracts — how robonix/prm/* and robonix/srv/* work
• Interface catalog — every primitive + service contract

Agent Skills (agentskills.io)

Robonix can ingest Agent Skills — the same SKILL.md+frontmatter format used by Claude Code, Cursor, GitHub Copilot, OpenCode, Gemini CLI, OpenClaw, etc. Place them under ~/.robonix/skills/<name>/ (or any directory in ROBONIX_SKILLS_EXTRA_DIRS); Pilot picks them up at startup.

Disambiguation: an "Agent Skill" here is an agent-context playbook — a Markdown file telling an LLM how to use a set of tools. It is not the same thing as the Skill layer in the Robonix whitepaper, which refers to deployable agent behaviors (basic skills = pretrained VLA/RL processes; RTDL skills = runtime-generated structured plans). The two share a name but live at different abstraction levels. See docs/src/skill-library.md for the Robonix Skill model — Agent Skills are just one possible source of SkillInfo injected into the VLM prompt.

Earlier versions auto-scanned <package>/skills/ and registered entries to Atlas. That mechanism has been removed — packages and skills are now decoupled. See the migration note in skill-library.md.

rbnx CLI

rbnx --help for the full list. Most-used:

rbnx setup                # register the clone so other packages find contracts/IDL
rbnx build -p <pkg>       # run the package's build.sh (codegen + whatever)
rbnx codegen -p <pkg>     # just the codegen part (proto + Python stubs)
rbnx start -p <pkg> -n <node>
rbnx chat                 # TUI → Pilot → Executor
rbnx nodes / tools / channels / inspect / graph

Package authors: see Build & Codegen.

Status

Warning

Early, fast-moving development. Interfaces, IDL layouts, internal designs may change without notice. No API stability until a versioned release.

Working today: Atlas control plane, Pilot/Executor ReAct loop, SKILL.md discovery, rbnx package flow (validate / build / codegen / start / install), Tiago Webots E2E, ManiSkill3 VLA demo.

In progress: formal system-service deployment (services are example packages today), TaskGraph beyond linear, catalog enforcement on the server, product-grade Liaison.

Contributing

Fork from dev, make fmt, make check, PR. See CONTRIBUTING.md if present.

License

Mulan Permissive Software License, Version 2 (MulanPSL-2.0). See LICENSE.