AI doesn't just reply with words — it turns toward you, comes closer, and wags its tail. Sirius is a robot dog built around the responses of its small body. What deserves attention is not just its dog-like appearance. The path from reading voice and imagery to converting them into physical movement reveals the design thinking of consumer AI hardware. Compared with Meydo S1, which is at the product-concept and user-needs validation stage, the question of "what exactly does AI move" becomes clearer.

Figure 1 An edited image placing a still frame from Sirius's official animation next to the Meydo S1 concept. Displayed proportions are not actual size and do not indicate S1 feature implementation.
A ~1kg Body as the Exit Point for Conversation
Sirius is Hengbot's AI robot dog, and the Japanese Makuake project is run by YMTREND. Official material introduces a body of about 1kg, movement of up to about 0.5 m/s, and motions such as walking, turning and jumping. The compact body concentrates 3 actuators per leg, feeding back the angle, angular velocity, load and temperature of each degree of freedom. To turn a language AI's answer into motion, this kind of low-level control is indispensable.
That is why, when looking at Sirius, it is important not to treat the conversational model and the movement mechanism as a single capability. However naturally it understands speech, the motions it can execute are determined by floor conditions, range of motion, center of gravity and the drive system. The official maximum speed is a value under specific usage conditions, and this article has not measured driving or fall avoidance. When used at home, steps and wet areas must be avoided — it is not a waterproof product.

Figure 2 A still frame from Sirius's official animation. It introduces the joint mechanism and is not a third-party measurement of driving performance.
Combining Nearby Perception with Cloud Understanding
The head is described as equipped with an 8MP camera, dual microphones, a ToF sensor and touch input, carrying a 5 TOPS AI chip. Inputs such as turning toward the direction it was called, sensing the distance to a person, and reacting to being petted connect to changes in expression and posture. The 1.8-inch display serves more as a face than as an information display, with the design conveying responses through the whole body.
The processing split stated explicitly by the manufacturer also matters. Face recognition runs on the device, and face data is not uploaded to the cloud. Meanwhile, conversation content and intent analysis use cloud AI. Utterances like "come here" or "do a dance" are converted into combinations of movement, sound and expression. Just because face recognition is local does not mean everything, including conversation, completes offline.

Figure 3 Official Sirius image. It describes on-device face recognition but does not mean all processing runs offline.
Separating the Fun of Teaching Motions from the Development Environment
In addition to specifying reactions in natural language, manual teaching is also introduced: guiding joints by hand, recording posture and force information, and reproducing it as a motion. This reduces the burden of programming from zero and serves as an entry point for users to change their robot's behavior. But it does not mean any specified motion can be executed. It should be understood as building expressions within controllable motions and mechanical limits.
Furthermore, the EDU edition offers APIs supporting ROS 2, C, C++ and Python, as well as Creator Studio Pro for editing motions and expressions. The October 23 official notice states that API access is enabled after individual contact and a reply with required information. Customization and the development environment for the standard edition are a separate offering. For research or production, check the edition and usage procedure. Extensible functionality is valuable, but it is not grounds for assuming household chores are handled autonomously the moment you buy it.

Figure 4 A still frame from Sirius's official animation. The production environment differs by edition; EDU-oriented features should not be treated as universal across models.
The Japanese project started on August 28, 2026 and is scheduled to end on October 30. General rewards are scheduled for delivery by the end of December. Meanwhile, an October 23 official report stated that the first batch for backers from the opening day had shipped from an overseas warehouse. The plan envisions arrival in Japan on October 26, about 2–3 days of customs clearance, domestic shipping from October 28, and delivery by November 8 at the latest — this is a plan, not a confirmation of arrival or completed delivery. Check the latest notice for your batch.
The page's figure of about ¥148 million is the overseas funding amount from the 2025 Kickstarter, not this Makuake total. More important than the size of the number is reading the experience and delivery timing it corresponds to.

Figure 5 Sirius's perception-to-response flow organized from official material. It is an edited diagram simplifying the processing split, not measured results for latency or recognition rates.
Meydo S1 Aims to Expand the Target of Action to the Screen
Where Sirius moves a robot's legs and expressions, the material for Meydo S1 sketches an entry point for recording everyday information, confirming it, and then proceeding to digital work. Its concept receives voice, images, camera footage and screen data, connecting to transcription, summaries, to-dos and personal memory. S1 is at the product-concept and user-needs validation stage, and it cannot be stated that these features are complete on real hardware.
For example, recording thoughts while out, then reviewing them at the desk together with related information. In a meeting, sorting out owners and deadlines from what was said, with the user confirming the content and choosing the next task. Where Sirius's motions return visible physical results, in this use case the records and task states are the result. Both are AIs responding to input, but where failure is judged differs.

Figure 6 A concept image of Meydo S1 used for everyday recording. It shows product appearance and usage intent, not proof of real-device operation or recognition performance.
S1 also has a plan for a USB KVM mode that connects to a smartphone or PC via USB. The design understands permitted screens and converts instructions into keyboard, mouse and touch operations. Just as robot development must consider the body's range of motion and floor conditions, here the OS, apps, authentication screens and connection conditions determine the scope of execution. Compatibility and execution accuracy are unverified at present; this is not yet the stage where plugging in USB completes arbitrary work.
The concept uses three physical buttons — AI, Confirm, Cancel — where the user approves operations presented by the AI and can stop them when needed. FoneClaw manages records, memory, permissions, sync and task state. The ability to later change or delete approved-and-saved information is also a premise of personal memory. The idea of connecting confirmation to execution to result reporting is concrete, but whether operations can be performed correctly on each screen needs to be verified going forward.

Figure 7 An edited diagram simplifying Meydo S1's confirmation-and-execution concept. Operation of external devices is unverified and does not indicate success rates or the implementation of a stop function.
Making "AI That Moves" Concrete
Sirius's value lies in responding with its body to voices and touch, and even offering the fun of creating motions oneself. Having a path from perception to action is an important characteristic of embodied AI. But that alone does not make it a general-purpose home agent that autonomously completes tidying, chores or multi-step errands. What can be verified in public materials should be separated from future expansion possibilities.

Figure 8 A comparison of Sirius's public information and Meydo S1's concept. It organizes the design that moves a body and the design of digital work, and does not indicate equivalence of product stage or performance.
If you want to enjoy interacting with a robot and making motions, Sirius's joint structure, input methods and EDU environment are your decision criteria. If you want to move from recording to work on a PC or smartphone, it is worth waiting to see what S1's concept can actually complete. S1's price, weight, battery life, connectivity and supported devices, release and shipping are undetermined. To evaluate AI action, one needs to specify concretely: what moves, where it stops, and what counts as completion.
