Galbot combines embodied models with mobile handling and developer tools

Galbot develops embodied AI models and the robots that use them for physical work. Its catalogue separates mobile manipulation, heavy handling and expressive humanoid interaction. The developer offer is unusually important to the buying decision: hardware, development tools, scenario kits and a complete deployed solution are distinct layers with different availability and commercial scope.

01 / ProductOne model programme serves several robot forms

Galbot’s company page describes AstraBrain as its embodied model programme and AstraData as the data infrastructure behind it. The site presents a combination of perception, decision making, manipulation and movement control. Treat claims of universal intelligence or unique global leadership as the company’s positioning; they are not necessary to understand the concrete products offered.

The G1 page describes a wheeled robot with two arms, a lifting body and replaceable end effectors. Its specification lists camera and lidar inputs and supports changing between hands, suction cups and grippers. This is a mobile manipulation platform: the body must reach a useful pose and the selected tool must then perform the object interaction.

S1 is positioned for industrial, warehouse and logistics handling. The vendor lists a maximum dual-arm payload of 50 kg and hot-swappable batteries. Those are configuration specifications, not evidence that every 50 kg object can be picked from any orientation or moved through any environment. Object shape, grip and load distribution remain part of the application.

ET1 follows a different route, emphasising real-time interaction and expressive movement for settings such as greeting and performance. It has standard, Pro and Flagship configurations. A reader should not transfer S1’s heavy-handling scope to ET1 or assume that ET1’s interaction features define the industrial platform.

02 / AudienceSelect the development layer as well as the robot

The developer portal separates three layers. Basic tools include an SDK, documentation, simulation assets and development services. Scenario kits package narrower applications. Complete solutions involve Galbot and partners delivering a customised operating result. The portal says the layers are separately charged and may be purchased individually or combined.

A robotics laboratory is likely to care about low-level observations, control and the ability to reproduce experiments. A systems integrator may instead want a useful starting application with configuration tools. An end customer may prefer one accountable delivery team. These are different needs; selecting the most sophisticated robot does not resolve which party performs the engineering.

A retailer considering item retrieval should describe the assortment before discussing general-purpose manipulation. Packaging, transparent surfaces, deformable bags and crowded shelves can create different perception and grasping problems. The useful commercial question is which portion of the assortment the configured system can handle under an agreed operating process.

A warehouse with heavier loads should start from the load and workspace rather than the headline payload. Establish how the object is presented, how it is gripped and where it is released. A handling system that works with one standard tote may need substantial additional engineering for a mixed stream of parcels.

03 / WorkflowA proposed greeting application uses the currently live kit

For a first developer project, consider a supervised visitor greeting that directs people to a staffed reception desk. This is a proposed example, not a Sequenced test. It uses the portal’s currently live greeting and guidance kit as its starting point, rather than assuming the planned inspection or forthcoming sorting packages are already available.

Define the approved knowledge and interaction style first. The robot should provide location information, respond to a limited set of questions and hand uncertain requests to staff. Keep business commitments and sensitive personal decisions outside the initial scope. A welcoming presentation should not encourage visitors to rely on unsupported answers.

Then map the interaction to the site. Establish where the robot stands, what movements are permitted and how it detects that a person has engaged. A gesture useful in an empty demonstration area may be confusing or obstructive near a busy doorway. Test the behavior from different approach directions and with ordinary background activity.

Use simulation for the aspects it can meaningfully represent, then validate on the delivered robot. The developer portal describes simulation assets and a path to deployment, but a simulated scene cannot prove speech recognition in the actual lobby or predict every visitor reaction. Keep the transfer between environments explicit in the project record.

Finally, evaluate successful visitor routing, misunderstood requests and staff interventions. Distinguish problems in knowledge, speech interaction and physical positioning. The next development step might be a better script or a different location, rather than a more complex model. This bounded result provides evidence for deciding whether broader navigation or manipulation would add useful value.

04 / PricingDeveloper subscriptions are separate from the robot purchase

On 3 October 2026, the Chinese-language developer subscription table listed Free at CNY0, Pro at CNY50,000 per robot per year and Enterprise at CNY200,000 per robot per year. Paid plans are annual contracts. Enterprise also states a multiple-robot purchase requirement and directs larger arrangements to commercial discussion. These are developer subscription figures, not hardware prices.

The Free tier includes SDK, documentation, digital robot models and community support. The portal places online simulation and platform deployment among paid benefits. A team should therefore verify which part of its intended workflow is included at its chosen tier rather than assuming that free access supplies every hosted service.

For hardware and a complete application, obtain a separate quote. The G1 and S1 product pages provide purchase routes, but this review did not verify an all-inclusive current hardware tariff through those external stores. A project budget should preserve the distinctions between the body, subscriptions, scenario kits, integration and ongoing operation.

05 / DistinctionsThe offer exposes the gap between research tools and finished work

Galbot’s layered catalogue is useful because it makes the delivery choice visible. A research team can discuss lower-level tools, an integrator can start from a scenario kit, and an operator can seek a complete solution. Those routes can reduce ambiguity about who owns the engineering, provided the contract follows the same boundaries.

The Dexterity blueprint provides context for industrial manipulation evaluated around a completed handling application. Compare the actual objects, throughput conditions and integration ownership. A broad embodied-model narrative does not replace evidence for the physical process a buyer wants to automate.

The Physical Intelligence blueprint focuses on the model layer of robotics. Galbot’s offer also includes bodies and development services. That difference helps a team decide whether it is selecting a model research direction, a hardware platform or a delivered operating system for a particular task.

06 / QuestionsRead the live, forthcoming and planned labels literally

The developer product map marks greeting and guidance as available, material sorting as coming soon and site inspection as planned. Those labels are consequential. The fact that a product overview discusses sorting or inspection does not mean a turnkey development kit can be purchased and deployed today. An integrator may still build a custom application, but that is a different commitment.

The S1 page also pairs continuous-operation language with an eight-hour full-battery runtime specification. Read the former as an energy-management proposition involving battery replacement, rather than a claim of infinite runtime from one charge. A site trial should observe the complete replenishment process and the people or infrastructure it needs.

For model generalisation, ask for examples close to the intended environment and keep the task boundary visible. A successful manipulation demonstration does not establish how the system handles unfamiliar packaging, a changed shelf layout or an unavailable object. The most valuable evidence explains both the ordinary success path and the behavior when the robot cannot proceed.

07 / DecisionBuy the level of engineering responsibility you need

Galbot is relevant to teams that want embodied AI linked to concrete robot hardware and a visible development route. Its current portal gives buyers more to assess than a general company mission: there are product layers, support tiers and explicit availability labels. A useful decision starts by choosing the task and deciding who will engineer it.