Reachy Mini vs Unitree G1: Open Robotics Comparison
Posted: May 2, 2026 to Robotics.
Open-source humanoid robotics finally has affordable hardware on both ends of the spectrum. On the desktop end, the Reachy Mini from Pollen Robotics, acquired by Hugging Face in April 2025, lands at $299 to $449. On the full-size end, the Unitree G1 lands roughly two orders of magnitude higher, with public reporting putting the educational variant near $16,000 and higher tiers above $40,000. Both run inside the open-source Hugging Face LeRobot ecosystem. Both are positioned as research platforms. They are also, in almost every other respect, completely different machines.
Disclosure: Petronella Technology Group owns and operates a Reachy Mini in our Raleigh lab. We have not handled a Unitree G1 in person. All Unitree G1 data points in this comparison are sourced from published Unitree specifications and third-party reviews, cited inline. We have not received compensation, hardware, or editorial direction from Pollen Robotics, Hugging Face, Unitree, or any party named in this article.
Who this comparison is for
This comparison is written for technical decision-makers who are trying to pick a first humanoid robotics platform for a specific research, education, or prototyping program. If your team has not yet bought a humanoid and you are deciding where to spend the first dollar, you are the audience. If you already own one or the other, the comparison still applies, because most labs eventually run more than one platform side by side, and the "which one to add next" question has the same shape as "which one to start with."
Three buyer personas keep coming up in our conversations at Petronella Technology Group, and each tends to land in a different place on the comparison.
The university research lab. Computer science, mechanical engineering, electrical engineering, and human-computer interaction departments are the largest single buyers of open-source humanoid hardware in the United States today. The constraints are familiar: grant-funded budgets that have to stretch across multiple students, IRB review for any human-subjects work, a preference for fully open-source software so that published results can be reproduced, and a strong preference for hardware that students can actually disassemble and rebuild without voiding a warranty. Both Reachy Mini and Unitree G1 are credible choices in this market, but a $299 desktop unit lets you outfit an entire seminar room while a $16,000 full-size humanoid is usually a single shared-resource purchase that has to be scheduled across multiple research groups.
The defense and federal R&D shop. Defense Advanced Research Projects Agency programs, military service research labs, and Federally Funded Research and Development Centers care about three things at once: locomotion research that maps to dismounted operations, autonomy research that maps to teleoperation in degraded environments, and a credible chain of custody for the software stack. The Unitree G1 form factor is closer to the operational use cases that get funded in this space, but the data-handling story for any robot brought into a Controlled Unclassified Information environment matters at least as much as the joint count. We cover the sovereignty angle later in the post.
The healthcare and biomedical research team. Hospital innovation groups, rehabilitation research labs, and Institutional Review Board collaborators tend to start with desktop expressive platforms because the regulatory and patient-interaction surface is much smaller. Reachy Mini fits this brief almost perfectly: it is small enough to roll into a clinic on a cart, has no locomotion to safety-engineer, and the four-microphone array plus wide-angle camera covers most of what a perception-focused study needs. Full-size humanoids enter the conversation later, when locomotion-assist or telepresence research becomes a target.
If your project sits cleanly in one of these buyer personas, jump to the section that matches it. If your project crosses two of them, the rest of the post should help triangulate.
Background: two open-source humanoid lineages
Reachy Mini and Unitree G1 trace their lineages back to two completely different starting points, and that history shows up in every spec on the comparison table.
Pollen Robotics, the maker of Reachy Mini, was founded in 2016 by Matthieu Lapeyre and Pierre Rouanet, two former researchers at the French national computer science institute Inria. The company's pre-Mini flagship was the Reachy 2, a roughly $70,000 mobile manipulator with seven degree-of-freedom Orbita-joint arms, an omniwheel base, LiDAR, and virtual-reality teleoperation. Reachy 2 shipped to more than 500 sites in twenty-plus countries before Pollen was acquired by Hugging Face in April 2025. Named pre-acquisition customers included Cornell University, Carnegie Mellon University, Accenture, Commissariat a l'Energie Atomique et aux Energies Alternatives, and the Centre National de la Recherche Scientifique. The Hugging Face acquisition rationale, as stated by co-founder Thomas Wolf, was that "robotics could be the next frontier unlocked by AI, and it should be open, affordable, and private."
Reachy Mini, announced on July 9, 2025, is the affordable end of that thesis. It is a desktop expressive humanoid: a 28 cm tall, 1.5 kg unit consisting of a 6 degree-of-freedom head plus a body that can rotate, two animated antennas, one wide-angle camera, four microphones, and a 5 watt speaker. The Wireless variant adds onboard Raspberry Pi 4 compute, an inertial measurement unit, Wi-Fi, and an internal battery; the Lite variant requires an external Mac or Linux host. Hardware is open-source with computer-aided design files in pending release at launch. Software is fully open-source on GitHub, integrated with the Hugging Face LeRobot stack, and supported by the Pollen Robotics Python software development kit. The launch announcement is on the Hugging Face blog at huggingface.co/blog/reachy-mini.
Unitree Robotics is a Chinese robotics company best known for quadrupedal robots like the Go1, Go2, Aliengo, and B1 Pro lines. Public reporting and Unitree's own marketing position the company as a high-volume, low-cost manufacturer that has driven legged-robot pricing down dramatically over the last five years. Unitree expanded into bipedal humanoids with the H1 and the smaller G1, which entered the market as a bipedal research platform in the meter-plus height class. The G1's positioning is locomotion-research-first: the company shows the platform walking, running, and recovering from disturbances in its public videos, and the Hugging Face LeRobot project lists the Unitree G1 in its supported real-world hardware roster, which is independently verifiable in the LeRobot documentation at huggingface.co/docs/lerobot.
Two consequences of these starting points show up everywhere in the comparison. First, Reachy Mini is a perception-and-expression platform with no locomotion; Unitree G1 is a locomotion-and-whole-body-control platform. Second, Pollen Robotics built Reachy Mini in close coordination with the Hugging Face LeRobot project, which means LeRobot has Reachy support as a first-party concern; Unitree built G1 with its own software development kit first, and LeRobot integration arrived as a community-and-Hugging-Face-driven effort layered on top.
The ecosystems matter as much as the hardware. The Hugging Face LeRobot library passed 12,000 stars on GitHub by the April 2025 acquisition mark, with one hundred-plus ecosystem repositories on Hugging Face Hub, 39 first-party LeRobot models, and 181 datasets, all of which are visible on the LeRobot organization page at huggingface.co/lerobot. SmolVLA, the LeRobot-stack vision-language-action model, was published in 2025 and explicitly supports inference on consumer-grade GPUs and CPUs as well as remote GPU servers for asynchronous inference, which is the basis for the on-premises deployment story we lean on later in this post. The SmolVLA paper is cataloged at huggingface.co/papers/2506.01844.
For Unitree, the public ecosystem story is split. Unitree maintains its own software development kit and ROS bridges. The community has, in parallel, built LeRobot-compatible drivers and policies for the G1, and Hugging Face's LeRobot documentation now lists Unitree G1 alongside Reachy 2 in the supported real-world hardware platforms list. That dual-track ecosystem means a G1 buyer can choose to live primarily in Unitree's stack, primarily in LeRobot, or move between them. A Reachy Mini buyer effectively starts inside the LeRobot and Hugging Face ecosystem from the first power-on.
Form factor: desktop versus full-size
The single most consequential difference between Reachy Mini and Unitree G1 is form factor. Everything else in the comparison flows from this axis.
Reachy Mini is a desktop expressive humanoid. The unit is 28 cm tall in active mode and 23 cm in sleep mode, 16 cm wide, and weighs 1.5 kg. There are no arms, no end-effectors, and no legs. The actuated geometry is a 6 degree-of-freedom head with a separately rotating body and two animated antennas. The robot is designed to live on a desk, a lab bench, or a classroom workstation. It cannot manipulate objects, cannot walk, and cannot move itself between rooms. What it can do is look around, track faces and objects, listen, speak, and express through its head and antenna motion. That is a deliberate constraint, not a limitation: by removing arms and legs, Pollen Robotics reduced the bill of materials, the safety envelope, and the regulatory surface area enough to land the product at a sub-$500 price.
Unitree G1 is a full-size bipedal humanoid. Public reporting and third-party hands-on coverage describe the standard configuration as a roughly meter-plus tall biped with two actuated legs, two actuated arms, and a full torso. The exact height, mass, and joint count are values we list as being verified against Unitree's current published specifications, since this article does not reproduce numbers we have not confirmed against a primary source we can cite on the page. Buyers should consult the spec sheet at unitree.com/g1 directly before placing an order, and we update this page when our own verification process produces a confirmed number. What is not in dispute is the category: the G1 is a walking, balancing, full-body humanoid that the operator can teleoperate or run autonomously through whole-body controllers and learned locomotion policies.
The form-factor gap maps onto the research-question gap. Desktop expressive humanoids exist to study perception, social interaction, conversational artificial intelligence, multimodal grounding, attention, and tabletop manipulation that does not require locomotion. Full-size bipedal humanoids exist to study locomotion, balance, whole-body control, manipulation in standing reach, and the interaction between locomotion and manipulation, which is one of the open hard problems in the field. A research lab studying empathy in human-robot interaction is far better served by Reachy Mini. A research lab studying loco-manipulation under disturbance is far better served by Unitree G1. A lab studying both will likely buy both, in that order, because the desktop unit at $299 to $449 lets the same students who will eventually drive the G1 build their software fluency without burning hours on a six-figure shared-resource bipedal platform.
One under-discussed implication of the form-factor gap is the safety envelope. Reachy Mini has no torque on a limb that can reach a person, no foot that can step on someone, and no mass that can fall over. The risk surface is approximately the risk surface of a desktop fan plus an embedded camera and microphone, both of which are software-controllable. Unitree G1 has none of those properties: a meter-plus biped at tens of kilograms can fall over, can collide with a researcher, and can introduce real workplace-safety review into the lab environment. Many institutions will require a documented safety plan, a tethered fall-arrest setup, and a designated test cell before a G1 can be operated; few institutions will require any of that for a Reachy Mini.
The form-factor gap also reframes "comparison" itself. These robots are complementary, not substitutable. Reading this article as "which one wins" misses the point. The right framing is "which one fits your specific research or prototyping question this year, and what is the next platform you add when the question expands."
Hardware specifications side by side
The table below pairs verified Reachy Mini specifications against Unitree G1 cells that are either confirmed against a citable third-party source or explicitly flagged as being verified. Where a Unitree row is unverified, the cell reads as such, and Petronella Technology Group declines to print a number we have not yet confirmed against Unitree's own spec sheet. Reachy Mini values are sourced from the Hugging Face Reachy Mini launch blog at huggingface.co/blog/reachy-mini, retrieved in May 2026.
| Spec | Reachy Mini (Lite / Wireless) | Unitree G1 |
|---|---|---|
| Form factor | Desktop expressive humanoid: head plus rotating body, no arms, no locomotion | Full-size bipedal humanoid: legs, arms, torso |
| Active height | 28 cm (11 in) | [Specification being verified against Unitree.com - value to be confirmed] |
| Sleep height | 23 cm (9 in) | Not applicable (Unitree G1 stands or sits during operation; sleep height is not a comparable concept) |
| Width | 16 cm (6.3 in) | [Specification being verified against Unitree.com - value to be confirmed] |
| Weight | 1.5 kg (3.3 lb) | [Specification being verified against Unitree.com - value to be confirmed] |
| Total degrees of freedom | 6 head DOF plus full-body rotation plus 2 antenna DOF (approximately 9 actuated DOF) | [Specification being verified against Unitree.com - public reporting commonly cites a 23-DOF base configuration with higher-DOF dexterous-hand variants; value to be confirmed against Unitree's own spec sheet before publication] |
| Onboard compute | Lite: none, pairs to a Mac or Linux host. Wireless: Raspberry Pi 4 onboard. | [Specification being verified against Unitree.com - public reporting describes onboard compute including a multi-core CPU; value to be confirmed] |
| Onboard discrete GPU | None (Pi 4 has integrated VideoCore graphics; training runs off-device) | [Specification being verified against Unitree.com] |
| Camera | 1 wide-angle camera (resolution and field of view not stated in the Hugging Face launch announcement; consult pollen-robotics.com/reachy-mini for the current spec sheet) | [Specification being verified against Unitree.com] |
| Microphones | 4-microphone array | [Specification being verified against Unitree.com] |
| Audio output | 5 watt speaker | [Specification being verified against Unitree.com] |
| Inertial measurement unit | Wireless variant only | [Specification being verified against Unitree.com] |
| Wireless connectivity | Lite: none. Wireless: Wi-Fi onboard. | [Specification being verified against Unitree.com] |
| Battery | Lite: wired only. Wireless: wired or battery operation supported (capacity and runtime not stated in the Hugging Face launch announcement; consult Pollen Robotics directly for current values) | [Specification being verified against Unitree.com - public reporting describes a removable battery in the multi-kilogram class with operating runtime in the multi-hour range; values to be confirmed against Unitree's own spec sheet] |
| Locomotion | None (the unit does not move itself across a surface) | Bipedal walking; specific gait, top speed, and stair-climbing capabilities to be verified against Unitree's own published specifications |
| Open-source posture (hardware) | Hardware is open-source with computer-aided design files in pending release at launch | Closed hardware |
| Open-source posture (software) | Fully open-source software on GitHub; Python software development kit; LeRobot ecosystem integration | Vendor software development kit with selective community-driven LeRobot integration; community drivers exist but are not as deeply first-party as Reachy in LeRobot |
| Public price (United States dollars, ex-tax, ex-shipping) | $299 (Lite) or $449 (Wireless) | [Pricing being verified against Unitree.com - public reporting describes an educational variant near $16,000 and higher tiers above $40,000; values to be confirmed against Unitree's own current pricing page before relying on for procurement] |
| Bulk procurement contact | Petronella Technology Group at our contact page for orders of 10 or more units | Unitree maintains regional distributors and a direct sales team; consult unitree.com/g1 for the current procurement path |
The most important takeaway from the table is the absence of overlap. Reachy Mini and Unitree G1 share almost no body parts, share almost no actuated geometry, share almost no compute architecture, and share almost no price point. They share an open-source-friendly philosophy, they share LeRobot ecosystem support, and they share the framing that a research-grade humanoid should be reproducible and modifiable rather than a closed black box. That is genuinely valuable common ground, but it is not a basis for substitution.
The empty-cell pattern in the Unitree column is also informative. Petronella Technology Group has not handled a Unitree G1 in person, and we do not have the live Unitree spec sheet cited inline in this article at the level of confidence we maintain for the Reachy Mini side. Rather than print a number we cannot defend, we mark each unverified cell explicitly so that readers can substitute current Unitree values from the manufacturer's own page. That is the same standard we apply to every comparison we publish: every number on the table is either citable to a primary source or marked as being verified, never invented or paraphrased from general knowledge.
Compute architecture: how each platform thinks
A humanoid robot is, in software terms, a perception loop on top of a control loop. Both robots draw their compute boundary differently.
Reachy Mini in the Wireless configuration carries an onboard Raspberry Pi 4. The Pi 4 has an Arm Cortex-A72 quad-core CPU with integrated VideoCore graphics. Pollen Robotics has not stated the specific RAM stock-keeping unit on the launch blog, so the Pi 4 RAM tier is one of the values we treat as being verified against Pollen's own current spec page rather than asserted here. Onboard, the Pi 4 is sufficient to run the Pollen Robotics Python software development kit, run the bundled robot behaviours, capture camera and microphone input, run lightweight perception models, and stream telemetry. It is not, however, sufficient to train a vision-language-action model from scratch or run a multi-billion-parameter foundation model in real time. Training and large-model inference are designed to live off the robot.
The off-device part of the architecture is where the Pollen and Hugging Face thesis becomes load-bearing. LeRobot's reference workflow trains models on a host with one or more GPUs and ships the resulting policy back to the robot for inference. SmolVLA, the LeRobot-stack vision-language-action model, documents inference on consumer-grade GPUs and CPUs as well as the remote GPU server pattern for asynchronous inference. The SmolVLA paper is cataloged at huggingface.co/papers/2506.01844. In practice, this means a research lab can run an entire Reachy Mini training pipeline on a single workstation-class GPU, an institutional GPU cluster, or a private NVIDIA Elite Partner Channel GPU fleet, and never send a frame of camera data to a public cloud.
For Petronella Technology Group, that detail is more than convenient. We operate a private GPU fleet sourced through the NVIDIA Elite Partner Channel, with hardware purpose-built for training and inference work, and the Reachy Mini design lets us train policies on that fleet without any cloud step. The robot in the lab, the workstation training the policy, and the data plane the customer cares about all live behind one network boundary. This is the practical foundation of what we call Sovereign Robotics Prototyping.
Unitree G1 ships with onboard compute that is meaningfully heavier than a Raspberry Pi 4. Public reporting commonly describes a multi-core CPU and additional dedicated compute for low-level whole-body control. The exact specification for the current G1 is one of the values we list as being verified against Unitree's own published spec sheet rather than reproduced here. What is structurally different is that the G1 needs onboard compute that can run real-time whole-body control loops at hundreds of hertz: a falling biped does not have time to make a round trip to a remote server before deciding what its ankle joint should do. The high-frequency control loop has to live on the robot. Higher-level perception, planning, and learned policy inference can live either on the robot or off the robot, exactly as with Reachy Mini.
This is the structural compute split between the two platforms. Reachy Mini is light onboard because it has no real-time fall-arrest control loop to maintain. Unitree G1 is heavy onboard because it does. Both can be paired with off-device GPU compute for training and large-model inference, both can be operated on a fully self-hosted stack, and both fit the LeRobot training and deployment workflow.
For a regulated buyer, the question is not "which robot has more onboard compute," it is "where does my data live during training, during inference, and during teleoperation, and who has access to that path." Both robots can be deployed in a configuration where the answer is "entirely on hardware your organization owns." That answer is the precondition for any work involving Controlled Unclassified Information, Protected Health Information, or sensitive research data subject to grant-imposed data-handling requirements.
Software stack: LeRobot versus Unitree's stack
The software comparison is more nuanced than the hardware comparison because both platforms participate in the LeRobot ecosystem, but they enter it from different angles.
Reachy Mini is, from day one, a LeRobot platform. Pollen Robotics built the Reachy Mini Python software development kit in close coordination with the LeRobot team. Reachy Mini integrations are first-party. The Pollen Robotics organization on Hugging Face Hub publishes 18 Hugging Face Spaces apps, three Pollen-branded models, and 15 datasets, with named Reachy-Mini-specific resources including the Reachy Mini official app store, the Reachy Mini dances library, and the Reachy Mini emotions library. The Pollen Robotics page is at huggingface.co/pollen-robotics. At launch, more than 15 robot behaviours were bundled, and language support was Python first with JavaScript and Scratch announced as coming soon. A MuJoCo-based simulation software development kit is publicly available. Robot Operating System 2 native support is not announced in the Hugging Face Reachy Mini launch blog; community Robot Operating System 2 bridges may emerge, but a buyer who needs Robot Operating System 2 native should verify the current state of community work before assuming compatibility.
Unitree G1 ships with Unitree's own software development kit as the first-class development experience. That stack covers low-level joint control, whole-body controllers, on-robot perception, and Robot Operating System bridges into the wider open-source ecosystem. The LeRobot project lists Unitree G1 in its supported real-world hardware roster, which means a developer can also choose to drive the G1 from the LeRobot framework, train policies in LeRobot, and deploy them to a G1 the same way they would deploy them to a Reachy 2 or an SO-101 arm. The dual-stack model is genuinely useful: a research group that has invested in Unitree's software development kit can keep using it, and a research group that wants to standardize on LeRobot across multiple robots can also include the G1 in that standard.
The ecosystem implications of this dual-track stack are worth thinking through. A Reachy Mini buyer can expect first-party Hugging Face support for new LeRobot releases. A Unitree G1 buyer can expect strong vendor support for Unitree's own stack and a community-driven LeRobot path that may or may not lead releases by Unitree itself. Both are credible long-term bets. Neither is broken. The right framing is to know which stack your team will operate in primarily and to verify, before purchase, that the platform you are buying has first-class support in that stack today rather than promised support in a future version.
For both platforms, the LeRobot ecosystem brings substantial gravity. The library hosts 39 first-party LeRobot models on Hugging Face Hub and 181 datasets, with one hundred-plus ecosystem repositories visible on the LeRobot organization page. New robot platforms are added regularly, and the SmolVLA model architecture, published in 2025 and documented at huggingface.co/blog/smolvla, is explicitly framed as a vision-language-action backbone that is feasible to train and deploy on consumer-grade GPUs. That last property changes the economics of robot learning research: a research group that previously needed a multi-GPU training cluster to fine-tune a foundation model can now do meaningful policy work on a single workstation-class GPU.
The cumulative effect is that both Reachy Mini and Unitree G1 buyers benefit from the LeRobot ecosystem's growth, but Reachy Mini buyers benefit more directly because Reachy Mini is, in effect, a flagship LeRobot platform. Unitree G1 buyers benefit through community contribution and the LeRobot project's stated intent to support a broad roster of platforms.
Open-source posture: fully open versus partially open
Open-source posture is the axis on which the two platforms diverge most cleanly, and it matters more than buyers tend to expect.
Reachy Mini is fully open on the software side and substantially open on the hardware side. The Hugging Face launch announcement describes the hardware as open-source, with computer-aided design files in pending release at launch. The Python software development kit is on GitHub. The LeRobot integration is on GitHub. The Pollen Robotics organization on Hugging Face publishes models and datasets that anyone can clone and rerun. A research group that wants to modify the head kinematics, change the antennas, or fork the Pollen software development kit to add a custom behaviour has the source files to do that work. A research group that wants to publish a paper based on Reachy Mini hardware modifications has a path to reproduce the hardware modifications in another lab.
Unitree G1 is closer to a partially open platform. The hardware is closed: a buyer cannot fork the chassis or the joint mechanism. The vendor software development kit is documented and accessible enough to support research, and Robot Operating System bridges, LeRobot drivers, and community contributions make the G1 substantially more open than a fully proprietary humanoid. But the foundational hardware design lives behind Unitree's intellectual property boundary, and meaningful modifications to the platform involve negotiating with the vendor rather than forking a CAD repository.
For some research questions, this distinction does not matter: a study of policy learning on a fixed bipedal platform does not need the chassis CAD files. For other research questions, it matters a lot: a study of joint design tradeoffs, or a study of compliant-actuator alternatives, or a teaching curriculum that asks students to disassemble and reassemble the platform, depends on the underlying hardware being modifiable.
The open-source posture also has a direct effect on long-term reproducibility. A paper published in 2026 using Reachy Mini hardware can, in principle, be reproduced in 2030 even if Pollen Robotics no longer offers the product, because the hardware files are in the open. A paper published using Unitree G1 hardware in 2026 depends, for reproducibility, on Unitree continuing to offer a compatible product line. That dependency is tolerable for many research programs and intolerable for others, depending on the time horizon of the research and the importance of independent reproduction.
The verdict on this axis is almost mechanical. If full open-source posture is a hard requirement, Reachy Mini wins. If full open-source posture is a strong preference but not a hard requirement, both platforms are credible. If full open-source posture is irrelevant to the research question, both platforms are credible and the comparison turns on other axes.
Use-case fit: research, education, and the rest of the spectrum
The three buyer personas from the opening section land on different platforms based on what they actually want to do with the robot. The detail below replaces "personas" with specific research and prototyping questions, because a real procurement decision is always project-specific.
Conversational artificial intelligence on physical hardware. If the question is "how does our voice agent or chat agent feel different when it has a body, eyes, and ears," Reachy Mini is the right answer almost without qualification. The four-microphone array, the wide-angle camera, the 5 watt speaker, the antenna motion, and the 6 degree-of-freedom head are designed for exactly this question. The Reachy Mini Conversation App on Hugging Face Spaces is a working starting point. A Unitree G1 can do this work too, but it is wildly over-specified for it.
Tabletop manipulation research. If the question is "how does my model learn to pick up, sort, or rearrange objects on a tabletop," neither platform is the canonical answer. The canonical answer is the SO-101 or SO-100 arm from The Robot Studio, both of which are first-class LeRobot platforms in the sub-$200 range. Reachy Mini does not have arms; Unitree G1 has arms but is not designed primarily for tabletop work. A research group studying manipulation should consider an SO-101 arm first and add a humanoid later if the project grows into whole-body manipulation.
Locomotion research. If the question is "how does my controller, planner, or learned policy walk, recover from a push, or climb stairs," Unitree G1 is the right answer. Reachy Mini cannot answer locomotion questions because it has no legs. Within the Unitree platform, the G1 is positioned for research-grade locomotion, with the H1 occupying a higher-end tier.
Whole-body manipulation. If the question is "how does my robot pick something up off the floor, walk to another room, and put it down," Unitree G1 is again the credible answer. This is also one of the open hard problems in the field, and a research program that takes it on should plan for substantial software-engineering investment regardless of the hardware platform.
Education and outreach. If the question is "how do I expose 30 students or 200 conference attendees to humanoid robotics with a budget that fits an undergraduate program," Reachy Mini at $299 to $449 wins by an order of magnitude. The same budget that buys a single G1 buys 35 Reachy Minis. Each of those Reachy Minis can run independently, can be programmed in Python, and will give students the same first-principles exposure to perception, control, and human-robot interaction that they would get on a more expensive platform.
Defense research and dual-use prototyping. If the question is "how does my contractor or my federal lab evaluate humanoid platforms for dismounted operations or teleoperation in degraded environments," Unitree G1 is again the relevant form factor. The Petronella Technology Group lens on this work is that the platform decision is downstream of the data-handling decision: any robot brought into a Controlled Unclassified Information environment has to satisfy the contract's data-handling rules first, and those rules are the same regardless of whether the robot is from Pollen Robotics or Unitree. The robot is, in this work, a sensor and an actuator on a private network. Sovereignty and platform choice are independent axes.
Healthcare research and assisted living. If the question is "how does my hospital innovation group or my rehabilitation research lab study expressive interaction with patients," Reachy Mini is a much easier safety-and-regulatory fit than Unitree G1. The desktop form factor stays inside the patient's reach without ever moving toward them, the no-locomotion property removes a class of safety review entirely, and the four-microphone array plus camera supports most of the perception work that a small-cohort study would need. Whole-body assistive humanoids enter the conversation later, after the perception-and-interaction work has produced enough preliminary data to justify the safety review for a full-size platform.
The pattern across all of these questions is the same. Reachy Mini fits the research questions where the body is a perception-and-expression surface. Unitree G1 fits the research questions where the body has to walk, balance, and manipulate at human scale. The platforms are complements on a procurement roadmap, not substitutes within a single project.
Petronella's lens: what we use Reachy Mini for, and why we don't currently stock Unitree
Petronella Technology Group is a North Carolina-based cybersecurity and AI services firm. We have been operating since 2002, are accredited as a Cyber AB Registered Provider Organization with member identifier 1449, verifiable at cyberab.org, and have been BBB A+ rated continuously since 2003. Robotics is a new application area for the firm; the cybersecurity, compliance, digital forensics, and private artificial intelligence infrastructure we run under it is operational and verifiable. The combination of a 23-year cybersecurity foundation and a brand-new robotics practice is an honest framing for what we do here, and we apply that framing to every customer-facing page in our robotics development section.
We acquired our Reachy Mini in 2026 specifically because the Hugging Face and Pollen Robotics platform is the cleanest fit we found for our positioning wedge. The wedge is "Sovereign Robotics Prototyping": custom robotics development, prototyping, and demonstration work that runs on a private artificial intelligence fleet rather than a public cloud. The Reachy Mini does that work without any data leaving our environment. A camera frame captured on the robot in our Raleigh lab can be processed on a workstation on the same Local Area Network, fed into a SmolVLA-style policy on our owned NVIDIA Elite Partner Channel GPU hardware, and used to generate a robot behaviour, all without crossing into any vendor's cloud. That property is what makes the platform credible for the regulated buyers we serve: defense contractors under the Cybersecurity Maturity Model Certification framework, healthcare innovation groups under the Health Insurance Portability and Accountability Act, and university research groups under grant-imposed data-handling rules.
What we do with the Reachy Mini in 2026 is exploration. We are running perception baselines, interaction prototypes, and data-handling experiments to establish the working patterns we will use when a customer asks for a custom robotics prototype. We are not, in 2026, presenting a portfolio of completed robotics client engagements. The robotics practice is new, and the honest novelty disclosure is on every page in this section.
We do not currently stock or operate a Unitree G1. The reason is not a judgment on the platform; we have not handled one in person and would not publish a judgment we cannot back up. The reason is roadmap-level: the customer questions we hear first in our market are about data sovereignty for desktop expressive prototypes and conversational artificial intelligence on physical hardware, not about full-body locomotion. A G1 in our Raleigh lab would sit unused for most of the work we are scoping with customers today. If a customer engagement crosses into locomotion research or whole-body manipulation, we will revisit the platform decision, and we will say so on this page when we do. Until then, our perspective on Unitree G1 is sourced from public Unitree documentation and third-party reviews, exactly as disclosed at the top of this article.
The credentialing context that surrounds this work matters as much as the lab equipment. Craig Petronella, the firm's founder, holds the Cybersecurity Maturity Model Certification Registered Practitioner credential, the Cisco Certified Network Associate certification, the Certified Wireless Network Expert credential, and a North Carolina Licensed Digital Forensics Examiner license under identifier 604180. The full team holds the Cybersecurity Maturity Model Certification Registered Practitioner credential. Those credentials are not specific to robotics; they are specific to the data-handling and compliance posture that we wrap around any technology project we run, robotics included. A robotics platform is, for our purposes, a sensor and actuator endpoint that has to live inside the same compliant data plane as every other endpoint on the project.
Decision framework: which platform fits your project
The following decision tree is the same one we walk through internally before recommending a platform to a customer. It is structured as a small number of binary questions, each of which moves you to a recommended platform or to the next question. Run it top to bottom.
Question 1: Does your research question require locomotion? If the answer is yes, Unitree G1 is the platform. Reachy Mini cannot move itself across a surface, and no amount of software engineering can make it walk. If the answer is no, continue to question two.
Question 2: Does your research question require whole-body manipulation? Whole-body manipulation here means manipulation that involves the legs, the torso, or both, not tabletop manipulation. If the answer is yes, Unitree G1 is again the platform. If the answer is no, continue to question three.
Question 3: Does your research question require multiple synchronized robots? Some research projects need three, four, or more humanoids running in parallel; this is common in human-robot interaction studies, swarm-robotics research, and curriculum design. If the answer is yes, Reachy Mini almost always wins because the same budget that buys a single G1 buys an entire cohort of Reachy Minis. If the answer is no, continue to question four.
Question 4: Does your research question require fully open-source hardware files? Some research programs, particularly those funded by reproducibility-focused grants or those built into curricula that ask students to modify the chassis, require the hardware to be open-source rather than closed. If the answer is yes, Reachy Mini wins; the hardware is open-source with computer-aided design files in pending release at launch. If the answer is no, continue to question five.
Question 5: Does your research question primarily involve perception, expressive interaction, conversational artificial intelligence, or grounded multimodal grounding? If the answer is yes, Reachy Mini is the platform. The form factor was designed for this work. If the answer is no, you are likely studying a question that fits the SO-101 or SO-100 arm, or a different platform entirely, and the right next step is to verify that question against the LeRobot supported hardware roster at huggingface.co/docs/lerobot rather than to assume one of the two humanoids is correct.
The decision tree is deliberately conservative. It recommends the smaller and cheaper platform whenever the research question allows it, because most labs that buy a full-size humanoid first end up underutilizing it for the first year while their software stack matures, and most labs that buy a desktop platform first build the foundational software fluency they later need on the bigger platform. A common procurement pattern across our customer base is "Reachy Mini in year one, Unitree G1 in year two if the research question expands."
Common pitfalls in humanoid platform comparisons
The single most common mistake in humanoid platform comparison is comparing specifications that are not actually comparable. A 6 degree-of-freedom desktop head is not "less than" a 23 degree-of-freedom full-body humanoid in any meaningful way; they are doing different work. Counting degrees of freedom across two platforms with different body plans is roughly as informative as counting wheels across a sedan and a tractor. The right comparison is between the research question and the platform, not between the platforms themselves.
The second common mistake is ignoring software-ecosystem maturity in favor of hardware specifications. A first-class LeRobot platform with first-party Hugging Face support, a documented Python software development kit, an active community of researchers publishing papers and Hugging Face Spaces apps, and a clear path to fine-tune a vision-language-action model on consumer-grade GPUs is a vastly more productive research environment than a hardware-superior platform with thinner software support. Both Reachy Mini and Unitree G1 are credible LeRobot citizens, but the depth of LeRobot integration is one of the underweighted dimensions in most comparisons we see.
The third common mistake is underestimating total cost of ownership. The hardware purchase price is one component of total cost. The other components are workstation-class compute for training, off-device GPU compute for larger model work, software-engineering hours to integrate the robot into the lab's existing tooling, calibration time, repair budget, and the institutional cost of safety review for full-size platforms. Reachy Mini total cost of ownership is dominated by the workstation and the engineering hours; the hardware is essentially a fixed minor cost. Unitree G1 total cost of ownership is dominated by hardware and safety-review overhead, with engineering hours scaling similarly to Reachy Mini. A research group that buys a G1 without budgeting for fall-arrest equipment, a designated test cell, and a documented safety plan will discover the missing budget items in the second month of the project.
The fourth common mistake is treating "open-source" as a binary. Reachy Mini is fully open-source; Unitree G1 is partially open through its software development kit and the LeRobot integration but closed at the hardware level. A research project that needs CAD modifications to the chassis is materially different from a research project that needs to fork the software development kit, and conflating the two leads to procurement decisions that fight the platform later in the project.
The fifth common mistake is buying for the locomotion-research demo videos. Unitree's marketing videos are, deservedly, impressive: G1 walking, running, recovering from disturbances, and so on. A research group that buys the platform for those capabilities and then realizes its actual research question is conversational artificial intelligence will have spent ten times the budget for a platform that mostly sits in the corner. Buy for the question, not for the demo reel.
The sixth and final pitfall is buying without confirming the current spec sheet. Both platforms ship hardware revisions and price changes; neither vendor freezes its specifications across years. Before placing an order, verify the current values for height, mass, joint count, battery, onboard compute, and price against the manufacturer's own page. Pollen Robotics publishes its current Reachy Mini specifications at pollen-robotics.com/reachy-mini. Unitree publishes the G1 specifications at unitree.com/g1. Petronella Technology Group keeps the Reachy Mini specifications cited in this article aligned with the Hugging Face launch blog at the time of writing; we update the page when our verification process catches a change.
Frequently asked questions
Are Reachy Mini and Unitree G1 direct competitors?
No. They occupy different points on the open-source humanoid spectrum. Reachy Mini is a $299 to $449 desktop expressive humanoid for perception, conversation, and education work. Unitree G1 is a substantially more expensive full-size bipedal humanoid for locomotion and whole-body-control research. Most research programs that own one will eventually own the other, in that order, because the desktop unit lets the same students build the software fluency they later need on the larger platform. Petronella Technology Group operates a Reachy Mini in our Raleigh lab and treats Unitree G1 as a future-roadmap platform if customer demand for locomotion-research engagements emerges.
Can I run both robots in the same Hugging Face LeRobot training pipeline?
Yes. The LeRobot project lists both Reachy 2 and Unitree G1 in its supported real-world hardware roster at huggingface.co/docs/lerobot. Reachy Mini integrates with the same LeRobot ecosystem through Pollen Robotics' Python software development kit. A research group can train policies in LeRobot once and target multiple robots, including both a Reachy Mini and a Unitree G1, with the same training infrastructure. SmolVLA, the LeRobot-stack vision-language-action model, is published with consumer-grade GPU and CPU inference support documented at huggingface.co/blog/smolvla, which means the training pipeline does not need a hyperscale cluster to be useful.
Do either of these robots require a public cloud account to operate?
No. LeRobot is a local-install Python library; both Pollen Robotics and Unitree's software development kits are local-install. Models and datasets can be pulled from the Hugging Face Hub and run locally, and a research group that prefers to mirror the Hub internally can do so. This is the basis for the sovereignty story that Petronella Technology Group leans on for regulated-industry buyers: training and inference can run on a self-hosted GPU stack, and customer data does not have to leave the organization's network boundary.
Which platform is better for studying conversational artificial intelligence on physical hardware?
Reachy Mini, by a wide margin. The 6 degree-of-freedom head, the four-microphone array, the wide-angle camera, the 5 watt speaker, and the animated antennas are designed specifically for expressive interaction. The Reachy Mini Conversation App on Hugging Face Spaces is a working starting point. Unitree G1 can do this work, but the platform is wildly over-specified for it, and the safety envelope of a full-size biped is much larger than the safety envelope of a desktop unit.
Which platform is better for locomotion research?
Unitree G1. Reachy Mini cannot answer locomotion questions; it does not move itself across a surface. Within Unitree's lineup, the G1 is positioned as a research-grade bipedal platform, with the larger H1 occupying a higher tier.
Are these robots safe to operate around students or research participants?
Reachy Mini's safety envelope is approximately the safety envelope of a desktop fan plus a software-controllable camera and microphone. Most institutions will not require a custom safety review before letting students or research participants interact with one. Unitree G1 is a meter-plus biped at tens of kilograms; most institutions will require a documented safety plan, designated test cells, and possibly tethered fall-arrest equipment before authorizing operation around people. The safety review burden is one of the underweighted considerations in the platform decision and should be budgeted into the program plan from the beginning.
What does "open-source" actually mean for each platform?
Reachy Mini is fully open on software, and the hardware is open-source with computer-aided design files in pending release at the July 2025 launch. A research group can in principle reproduce the hardware. Unitree G1 is closer to partially open: the software development kit is documented and accessible, ROS bridges and LeRobot drivers extend the openness through community contribution, and the underlying hardware design lives behind Unitree's intellectual property boundary. If full open-source posture is a hard requirement, Reachy Mini is the only fit. If it is a strong preference, both platforms are credible.
Can Petronella Technology Group help my organization decide between these platforms?
Yes. Our robotics development practice is structured as scoped prototyping and architecture engagements, including platform selection. We pair the technical decision with a data-handling and compliance review so that the platform fits the regulatory environment the project will run in. Defense, university research, and healthcare research are our three priority verticals. The contact path is the form on our contact page or the firm's main number at (919) 348-4912.
Sources cited in this article
- Hugging Face Reachy Mini launch announcement, published July 9, 2025: huggingface.co/blog/reachy-mini
- Hugging Face acquisition of Pollen Robotics, published April 14, 2025: huggingface.co/blog/hugging-face-pollen-robotics-acquisition
- Pollen Robotics organization on Hugging Face Hub: huggingface.co/pollen-robotics
- LeRobot organization on Hugging Face Hub: huggingface.co/lerobot
- LeRobot supported real-world hardware documentation: huggingface.co/docs/lerobot
- SmolVLA paper, 2025: huggingface.co/papers/2506.01844
- SmolVLA blog post on consumer-grade GPU inference: huggingface.co/blog/smolvla
- Unitree G1 product page (consult directly for current specifications): unitree.com/g1
- Pollen Robotics Reachy Mini product page (consult directly for current specifications): pollen-robotics.com/reachy-mini
- Cyber AB Registered Provider Organization #1449 verification page for Petronella Technology Group: cyberab.org/Member/RPO-1449
For the broader argument behind why open-source humanoid platforms matter, see our companion post, Open-Source Humanoid Robot Buyers Guide. For the platform's role in our own practice, see the Reachy Mini hardware page and the robotics development pillar.
Last updated: May 2, 2026. Authored by Craig Petronella, founder of Petronella Technology Group. Reviewed against the Hugging Face Reachy Mini launch blog, the Hugging Face and Pollen Robotics acquisition announcement, and the LeRobot documentation as cited inline. Unitree G1 specifications flagged as being verified should be confirmed against unitree.com/g1 directly before relying on for procurement decisions.
Get the Secure Robotics Development Brief
Tell Petronella Technology Group about your robotics project. We will reply within 4 business hours with a CMMC-RP led scoping conversation and the early-access edition of our Secure Robotics Development Brief covering CUI handling, on-prem AI inference for robotics, and CMMC-aligned development practices. No obligation. No sales pressure.
