Reachy Mini,
desktop robotics

Reachy Mini hardware specifications

The table below collects every published specification from the Hugging Face Reachy Mini launch announcement on July 9, 2025, retrieved May 2, 2026. Where a specification was not stated in the launch announcement, it is marked as being confirmed against Pollen Robotics documentation rather than guessed. Reachy Mini is open hardware: the software side is fully open-source and on GitHub today, and the CAD files were described as pending release at the launch announcement. We will update this page as additional values are confirmed against the Pollen Robotics product page.

A few things in that table are worth pulling out for engineering audiences. The 9-degree-of-freedom mechanical design (six in the head, two in the antennas, one in the body rotation) is a deliberate compromise: enough articulation for credible expressive interaction and head-motion-driven research, none of the safety, calibration, and certification burden that arms or legs would impose. This is what makes Reachy Mini useful as a teaching and research platform in environments, including classrooms and shared labs, where a full-body humanoid would not be welcome. The Wireless variant's onboard Pi 4 is sufficient to run sensor capture, behaviour playback, and small inference workloads, but heavier policy training and large vision-language-action models run on a paired host or, in our case, on Petronella's owned datacenter GPU fleet. The data-privacy default ("no personal data stored, transmitted, or processed by default") is the language we wanted to see from a vendor that lands on a desk inside a regulated-industry client. Specifications being confirmed against the Pollen Robotics product page or repository documentation will be updated on this page as soon as we publish those confirmations.

What we run Reachy Mini on, and why it matters

Reachy Mini's onboard Pi 4 (Wireless variant) is sized for sensor capture, behaviour playback, and small inference. Anything heavier than that, including policy training, vision-language-action model fine-tuning, and large-scale teleoperation data ingest, runs off-robot on owned compute. For the regulated buyers we serve, off-robot does not mean public cloud. It means our owned datacenter GPU fleet inside our managed infrastructure footprint, which is the same fleet that runs the rest of our private AI workloads.

Owned training inventory

We train robotics policies, fine-tune small vision-language-action models, and run large data conversions on a tiered GPU fleet sourced through the NVIDIA Elite Partner Channel. The fleet spans NVIDIA DGX systems for the heaviest workloads, NVIDIA HGX-class servers for production inference, and NVIDIA RTX PRO 6000 Blackwell workstations for development, prototyping, and small-scale training. The Reachy Mini integration, in practice, looks like this: data captured on the robot in our Raleigh lab streams to a development workstation, gets cleaned and converted to LeRobot's dataset format, and is then either trained locally on the workstation tier for small policies or shipped over our internal network to a DGX or HGX node for larger jobs. Inference for the resulting policy can run on the robot's Pi 4, on a Linux host on the same desk, or on a server-class node depending on latency budget. The LeRobot SmolVLA paper documents the same pattern explicitly: "consumer-grade GPUs or CPUs" for inference, with a "remote GPU server for asynchronous inference" pattern as a supported deployment.

If you are evaluating which compute to bring to a robotics prototype, our enterprise AI workstations page documents the same RTX PRO 6000 Blackwell, Threadripper PRO, Xeon W, and EPYC platforms that we pair with Reachy Mini for training-side work. Larger projects, including ones that need multi-GPU NVLink for vision-language-action model training, route through our private AI infrastructure stack.

Onboard inference path

Wireless Reachy Mini runs onboard inference on the Raspberry Pi 4. That budget is enough for the bundled 15-plus behaviours, for keyword-driven conversation flows, and for distilled small models compiled for ARM. When a workload exceeds the Pi 4's envelope, we fall back to the asynchronous-remote-inference pattern: the robot streams sensor frames over the local network to a paired Linux host or workstation, the host runs the policy, and the action stream returns over the same channel. We keep all of that traffic inside the client's network, the client's lab segment, or our own controlled enclave. Public-internet inference is not part of our default architecture.

Why owned, not cloud

The reason this matters, and the reason it is the wedge that earns us the right to do this work for regulated-industry clients, is that public-cloud robotics is a non-starter in the markets we serve. A defense contractor handling Controlled Unclassified Information cannot stream Reachy Mini sensor data into a tenant they do not own. A university research group with a National Institutes of Health-funded protocol cannot push human-subject video into a third-party model API without rewriting the protocol. A healthcare research group with Health Insurance Portability and Accountability Act-aligned data on the bench cannot route it through a vendor that has not signed a Business Associate Agreement. Owned compute, on premises, with a known provenance chain, is what makes a robotics prototype shippable in those rooms. That is what we built our cybersecurity practice on for twenty-three years, and it is the foundation we are extending into the robotics overlay now.

Reachy Mini vs the alternatives we considered

Reachy Mini is one of three open-source-leaning humanoid platforms a research team will reasonably evaluate today, and one of several research-grade platforms in the LeRobot ecosystem. The table below sets the comparison honestly. The two platforms are not direct substitutes. Reachy Mini is a $299 to $449 desktop expressive head-and-torso unit. The Unitree G1 is a full-size bipedal humanoid in the $16,000-and-up range. Hello Robot's Stretch 3 is a research-grade mobile manipulator. The three platforms answer three different questions, and the right answer for a given client depends on the question being asked.

The honest framing for a buyer: a $299 Reachy Mini Lite gets you onto LeRobot, lets you run perception and conversation experiments at low cost, and earns you a real entry into the open robotics ecosystem. It does not give you locomotion or arm manipulation. A Unitree G1 gives you locomotion and arm research at a $16,000-and-up price point with a different open-source posture. A Hello Robot Stretch 3 gives you mobile manipulation in a more clinical, ROS-2-native form factor. We chose Reachy Mini because it pairs the lowest cost-of-entry into LeRobot with a desk-friendly footprint that fits the early scoping conversations we have with the buyers we serve into. If a client needs locomotion or arm reach, we will scope the engagement to the platform that fits, not force-fit Reachy Mini into a job it cannot do. Specifications being confirmed against each vendor's published documentation will be updated on this page as confirmed; we link the canonical sources rather than reciting general-knowledge approximations as fact.

How to engage us on a Reachy Mini prototype

If you are reading this page because your team needs a robotics prototype on Reachy Mini hardware, with a Petronella security and compliance overlay, here is the path. We do not publish fixed pricing because the work is custom-quote by definition. The phases below are the same scoping pattern we run on every engagement.

1. Discovery call

   Thirty to forty-five minutes. Tell us the use case, the regulatory context (CMMC level, HIPAA, IRB, ITAR awareness), the rough timeline, and who the decision-makers are. We tell you what we have learned with Reachy Mini so far and where the platform fits, or does not fit, your problem. Outcome: a written one-page scope memo.

2. Architecture sketch

   One to two weeks. We produce a written architecture sketch that names the hardware (Reachy Mini Lite or Wireless plus optional LeRobot-supported peripherals), the compute (your fleet, our fleet, or hybrid), the data plane, the security overlay, and the deliverables. Outcome: a fixed-fee scoping document the right people on your side can sign.

3. Prototype build

   Four to ten weeks, scoped to your problem. We build, in our Raleigh lab on our owned compute and the Reachy Mini we already operate, with weekly demo touch-points back to your team. We do not invoice for shadow work; everything we build is visible.

4. Handoff

   One to two weeks. We deliver a hardware list (with the path to procure it through the Pollen Robotics direct channel for ten-or-more-unit orders), the code repository on your infrastructure, a written runbook, the security plan, and a recorded walkthrough. Optionally, we extend into a managed-prototype phase where we keep operating the prototype while your team learns the codebase.

The fastest path into the conversation is a phone call to (919) 348-4912 or a written scoping request through our contact form. We respond within one business hour during weekday business hours. We are not a high-pressure sales shop; if Reachy Mini is not the right platform for your problem, we will say so on the discovery call rather than after a contract.

Frequently asked questions