RoboParty RP1: the 100% open-source humanoid that opens hardware and software to researchers
🔎 Open-source finally takes on humanoid hardware
While open-source language models keep racking up breakthroughs — DeepSeek, Kimi, GLM — robotics hardware has remained a world of black boxes. You buy a research humanoid, you get a chassis, a proprietary SDK, and a support form. Electronic schematics? Nowhere to be found. The BOM? Classified.
On October 15, 2026, the Shanghai-based startup RoboParty decided to break this mold. It officially launched RP1 (ROBOTO 01), a full-stack open-source bipedal humanoid: mechanical design, electronic schematics, PCB layouts, firmware, control systems, and software — everything is published, viewable, and modifiable by researchers. The robot was first unveiled on September 28 at IROS 2026, in Pittsburgh.
Why now? Because RoboParty has already tested the formula. Its predecessor RPO, opened up in January 2026, has surpassed 2,600 GitHub stars and has rallied a community of more than 5,000 developers. The question is no longer whether open-source can extend to humanoid hardware, but whether openness alone is enough to dislodge the incumbents — Unitree at the top of the list — who have the pricing, the volume, and years of head start.
Key Takeaways
- RP1 (ROBOTO 01): full-stack open-source bipedal humanoid launched on October 15, 2026 by RoboParty (Shanghai, founded in 2025), following an unveiling at IROS 2026 on September 28.
- Two configurations: RP1 Standard (~1.3 m, 24 active degrees of freedom) and RP1 Pro (30 DoF, 2-axis neck, 7-axis arm, articulated RP Hand, interchangeable head modules and wrist effectors).
- Everything is published: mechanical CAD, schematics, PCB routing, Gerber files, BOM, firmware, drivers, ROS 2, Python SDK, training pipelines based on Isaac Sim 5.1.0 / Isaac Lab, and PartyOS, the embodied AI system.
- In-house actuators: Romomo, peak torque up to 160 N·m — a vertical integration choice that raises questions.
- Timeline: detailed open-source roadmap in October 2026, production program in Q4 2026. No price announced yet.
Recommended tools
| Tool | Main use | Price (October 2026) | Ideal for |
|---|---|---|---|
| GitHub Roboparty/RP1 | Official repos: hardware, firmware, deployment, training | Free | Getting CAD, BOM, schematics and code |
| ROBOTO 01 Documentation | Complete architecture, layer by layer | Free | Understanding the stack before buying |
| Isaac Lab (NVIDIA) | Training in simulation, Sim2Sim verification | Free (open source) | Training walking policies |
| ROS 2 | Robotics middleware, sensor and actuator integration | Free (Apache 2.0) | Prototyping your own modules |
| Hostinger | Hosting the docs and website of your open-source robotics project | From ~$3/month (check hostinger.com) | Publishing your results and attracting contributors |
| Compatible VR headset | Teleoperation and demonstration collection | Varies by model | Data collection for imitation (MimicLite) |
What is RP1, exactly?
RP1 is a bipedal research humanoid about 1.3 m tall, offered in two versions, with every technical layer — mechanical, electronics, firmware, control, AI — published on GitHub and officially documented.
The Standard version covers the essentials: 24 active degrees of freedom, walking, balance, and a sim-to-real pipeline. The Pro version goes up to 30 DoF, with a 2-axis neck, 7-axis arms (excluding hand DoF) and the articulated RP Hand. Head modules and wrist end-effectors are interchangeable, and compute is expandable — a rare feature in this segment, where most platforms lock you into their in-house machine.
| Version | Height | Active DoF | Key points |
|---|---|---|---|
| RP1 Standard | ~1.3 m | 24 | Walking, balance, sim-to-real |
| RP1 Pro | ~1.3 m | 30 | 2-axis neck, 7-axis arms, RP Hand, interchangeable modules |
| RPO (predecessor, 2026) | ~1.25 m | 23 | ~34 kg, open licenses, 2,600+ GitHub stars |
On the muscle side, RoboParty made a defining choice: Romomo actuators developed in-house, with peak torque reaching up to 160 N·m at the joints. We'll come back to this — it's both the project's technical strength and its main point of fragility.
"Full-stack open-source": what RoboParty actually publishes
This isn't open-washing: the published repos cover the entire chain, from mechanical design to physical deployment.
The RP1 repo on GitHub — presented as "An OpenSource Full-Stack Platform for Agentic Humanoids" — is broken down into documented modules:
- RP1_hardware: mechanical CAD, plans, electronic schematics, PCB routing, Gerber files, and a complete BOM.
- RP1_firmware: the embedded code for actuators and boards.
- RP1_deploy: the tools for physical setup.
- RP1_train: the training pipelines.
The software layer includes a Linux BSP, a USB-CAN-FD bridge, drivers for motors, IMU, hand, and cameras, a ROS 2 stack, and a Python SDK. The "intelligence" layer, for its part, exposes walking training, AMP, BeyondMimic, and Sim2Sim verification, according to the official documentation.
My take: the real move is the publication of the BOM and Gerber files. Any lab can audit the machine, plan its repairs, source its components, and — above all — verify what the robot actually does. That's exactly what established vendors refuse to provide.
A local clone takes one line:
git clone https://github.com/Roboparty/RP1.git
As for licenses, the previous RPO set the tone: CERN-OHL-W-2.0 for hardware, GPL-3.0 for software, BSD-3-Clause for training code. Three licenses, three levels of obligations — we'll come back to this in the common mistakes section.
The toolchain: from Isaac Lab to the real robot
RoboParty isn't just publishing a robot: it's publishing the complete chain from data collection to experimental validation, and that may be its most underrated contribution.
The pipeline covers four stages: data collection, simulation training, physical deployment, experimental validation. Concretely: VR teleoperation to collect demonstrations, retargeting to transfer them, training under Isaac Sim 5.1.0 and Isaac Lab, then whole-body control (WBC) for real-world execution. The UFO framework handles unsupervised RL, MimicLite handles motion imitation.
This sim-first approach isn't arriving in a vacuum. The sector is standardizing simulation training before any contact with the real world, and research on world models is rapidly taking shape — see our analysis of RoboJePa, the first scaling laws for robot world models published by Meta and Mila.
Why Sim2Sim verification is a game changer
The detail that matters for a lab: RP1 natively integrates comparison between model output and the robot's real-world behavior. You can therefore quantify the gap between what your policy does in simulation and what it does on the machine.
That's the foundation of reproducible science, not just YouTube demos. And it's a compelling argument against closed platforms, where this kind of verification remains an insider's hack.
PartyOS and the brain: which AI models for an open robot?
PartyOS, RoboParty's embodied AI system, bridges the gap between the open body and reasoning models — and that's where the choice of AI model becomes an architectural decision in its own right.
With onboard computing being extensible, teams have two strategies available. The first: a cloud brain, with the models that dominate today's agentic leaderboards — OpenAI's GPT-5.5 (98.2), Google's Gemini 3 Pro Deep Think (95.4), or Anthropic's Claude Opus 4.7 (94.3). Maximum power, but network dependency and per-request cost.
The second: local inference, more consistent with the project's philosophy. Self-hosted models like Moonshot AI's Kimi K2.6 (88.1 in agentic) or Z.AI's GLM-5 Reasoning (82) run on onboard hardware, and our guide on open source AI agents with Ollama running locally details how to orchestrate this kind of stack without sending a single frame to the cloud. Recent open-source edge models — such as iFlytek's Spark X2.5 and its two edge variants with 1M-token context windows — show that long context is no longer the exclusive domain of the datacenter.
The underlying trajectory? World models capable of piloting entire fleets of robots: Sber's GreenWAM already drives 20 robot families with a single model. A robot with an open-source body deserves a brain you can inspect — the equation holds.
Unitree Under Pressure: What RP1 Changes for the Market
RP1 doesn't claim to be the best humanoid on the market. It's selling something else: a new definition of what a lab is entitled to demand from a supplier.
The landscape remains dominated by closed platforms. Unitree and the other incumbents deliver remarkable robots, but without schematics, without a BOM, with opaque firmware and an SDK that remains their property. When a servo breaks out of warranty, the lab negotiates, waits, pays. With RP1, it opens the repo, reads the BOM, and orders the part.
This movement is part of a broader wave coming out of China, where open-source has become a weapon of commercial conquest. ByteDance has opened DeerFlow, its open-source agent that researches, codes, and creates over the long term, and OpenSeeker-v2 breaks the monopoly of industrial search agents. The sequence is always the same: where you can't win on marketing budget, you win on governance.
My prediction: if RP1 finds its audience, the incumbents' response will be cosmetic open-source — an open SDK, a published URDF, but neither schematics nor Gerbers. The frontier to watch, then, is not "open or not," but "how deep into the stack." RoboParty's slogan — "No More Black Boxes" — has the merit of setting the bar exactly where it needs to be.
The limits: openness doesn't build an ecosystem
Publishing files is the easy part. Price, reliability, and governance will decide what happens next — and on these three points, nothing is settled yet.
First point: price. No pricing has been announced to date (October 2026), with the production program slated for Q4. Yet RP1's entire value proposition rests on openness; if the bill ends up matching closed platforms, the only remaining argument will be transparency — solid for a lab, less so for a tight budget.
Second point, the most interesting one: the actuators. The RPO had a "repairable" BOM — more than 60% of its parts were off-the-shelf motors, available from any distributor. RP1 switches to in-house Romomo. RoboParty gains in performance and control over its supply chain, but creates a dependency on a single supplier… which is itself. The irony of an open-source robot relying on a proprietary actuator should be lost on no one.
Third point: openness is a process, not a state. The release is progressive — motion control, simulation, SDK, training tools, PartyOS — with a detailed roadmap announced for October 2026. Until everything is merged, "full-stack open-source" remains a roadmap promise.
Finally, who governs? PR merge policy, support, warranty: RoboParty hasn't detailed these mechanisms yet. They are what will turn 5,000 developers into an ecosystem — or into a mere audience.
Should You Equip Your Lab with an RP1?
For a lab that wants to modify the hardware, audit the full stack, and publish reproducible results, RP1 is probably the most coherent platform on the market — provided you accept buying early, into a young ecosystem.
What openness concretely brings: sourcing components yourself via the BOM, modifying the CAD for a specific end-effector, checking the firmware line by line, replaying the training pipelines, and quantifying the sim-to-real gap. No closed platform offers this level of control. The context confirms it: embodied AI research is moving fast, and the sector is closely followed by specialized publications like The AI Insider and Robotics International.
What openness does not bring: maturity. A community of 5,000 developers is promising but recent; no long-term field feedback on the Romomo; no price list; release roadmap still incomplete. My advice: wait for the detailed October roadmap and the first deployment feedback from Q4 before committing a budget — but bookmark the repos now.
❌ Common Mistakes
Mistake 1: believing that "open-source" means "reproducible at home"
A CAD, Gerber files, and a BOM are not enough to assemble a 1.3 m humanoid with actuators delivering 160 N·m. You need machining, power electronics, and integration experience. The solution: start in simulation with RP1_train and Isaac Lab, join the community, and leave physical assembly to well-equipped labs.
Mistake 2: treating the three licenses as a single block
CERN-OHL-W-2.0, GPL-3.0, and BSD-3-Clause impose very different obligations: the CERN hardware license requires sharing hardware modifications, the GPL contaminates your software forks, BSD-3 barely does. Publishing a commercial derivative without having read the licenses is the fastest way to get torn apart in public. Read them before your first contribution.
Mistake 3: skipping Sim2Sim verification
A policy trained under Isaac Lab can be excellent in simulation and mediocre on the real robot. If you benchmark without quantifying the gap between model output and real-world behavior, your results won't survive peer review. Use the provided verification pipeline and systematically log both sets of measurements.
❓ Frequently Asked Questions
How much does the RP1 cost?
No price has been announced to date (October 2026). RoboParty has announced a large-scale production program for Q4 2026; pricing should be communicated at that time. Keep an eye on the official channels and compare the total cost of ownership — parts via the BOM, tools, integration time — rather than the purchase price alone.
What's the difference between RP1 and RPO?
RPO (ROBOTO ORIGIN), released in January 2026, measured ~1.25 m for ~34 kg and 23 DoF, with a BOM made up of more than 60% off-the-shelf motors. RP1 goes up to 24-30 DoF depending on the version, switches to in-house Romomo actuators (up to 160 N·m), and publishes a more complete stack, PartyOS included.
Can you train your own walking policies?
Yes. The RP1_train repo exposes walking training, AMP, and BeyondMimic, based on Isaac Sim 5.1.0 and Isaac Lab. Plan for an RTX-class GPU for simulation, then validate with Sim2Sim verification before any physical deployment on the robot.
Can the RP1 do manipulation?
Yes, especially in the Pro version: 7-axis arm, articulated RP Hand, interchangeable wrist end effectors. VR teleoperation and retargeting make it possible to collect demonstrations, which MimicLite turns into imitation policies. It remains a research platform: don't expect to use it in industrial production.
Can you plug your own AI model into PartyOS?
PartyOS is designed as an extensible embodied AI system, with a Python SDK and extensible onboard computing. You can connect self-hosted models like Kimi K2.6 or GLM-5, or cloud APIs like GPT-5.5 — it's up to you to weigh latency, cost, and data privacy for your experiments.
✅ Conclusion
With RP1, RoboParty isn't just selling a humanoid: it's setting a new standard — a research robot whose every layer can be read, modified, and audited — and it's now up to the incumbents to prove why their black boxes still deserve our budgets. Start with the official GitHub repo and the ROBOTO 01 documentation: it's free, and it gives you a head start on everyone who will wait for the price.