Norwegian startup Kinetic Blocks has launched a marketplace for robotics training data, targeting developers of humanoid robots and other physical AI systems. The platform brings together datasets from different suppliers and adds a verification and grading system intended to make it easier for robotics companies to assess data before licensing it. Kinetic Blocks opened the platform in gated beta on 1 September.
Training data has become an increasingly important part of humanoid robot development. While large language models can draw on enormous quantities of text and images available online, robots require data about interactions with the physical world. This can include video of people carrying out tasks, demonstrations recorded through teleoperation and data generated by robots performing tasks themselves.
Kinetic Blocks wants to create a commercial marketplace for this type of data. Suppliers can list datasets while retaining ownership and determining the price, licensing conditions and whether access is exclusive. Buyers can search the marketplace based on characteristics including task, environment, modality and method of data collection.
The marketplace currently includes data captured in household environments, while additional datasets covering office work, workshops, personal care, industrial assembly, hospitality and construction are being prepared or tested. The company also offers to source datasets that are not already available through the marketplace.
Verification before sale
A central part of the Kinetic Blocks model is the verification of datasets before they are made available. Sellers provide a dataset card containing information such as the task, environment, robot, modalities, frame rate, resolution and licence. Kinetic Blocks then compares those claims with the underlying files.
Verification is separated from the company’s quality assessment. Each dataset receives a KB Quality grade ranging from A+ to F. The score is based on four areas: technical quality, content usefulness, temporal integrity and completeness. Kinetic Blocks says it also measures inactive periods in recordings rather than removing them, allowing buyers to assess how much of a dataset contains useful activity.
The first publicly visible listings illustrate the model. One consists of 20 hours of egocentric video showing household manipulation tasks, divided over 198 clips and priced at $800. Another contains ten hours of cleaning-related first-person data and is listed for $300. Both currently carry an A- quality grade.
The platform accepts conventional video as well as teleoperation and multimodal robot data. Kinetic Blocks recommends the LeRobot format for structured robot datasets but says suppliers can also use their existing data formats.
Links with humanoid robotics
Kinetic Blocks was founded in Oslo by a team with connections to the emerging humanoid robotics sector. CEO Lars-Fredrik Forberg previously worked as CTO of Scandinavian construction group Mestergruppen. Chairman Erik Ålgård spent around a decade on the board of Norwegian humanoid developer 1X Technologies, while co-founder Christian Rokseth was an early investor in 1X and also founded robotics publication Humanoid.guide. Leo Terry leads data quality and developed the verification and grading system.
The company operates from Startuplab in Oslo. Dotspot.ai carried out an initial distributed data-collection pilot, while Kinetic Blocks is also a member of NVIDIA’s Inception programme for startups.
Growing demand for physical data
The launch comes as humanoid developers are looking for ways to expand the amount and variety of data available for training robots. Teleoperation remains an important method: a human operator controls a robot while the system records observations and actions that can later be used for model training. But collecting large volumes of this data requires robots, operators, infrastructure and time.
Egocentric video offers another potential source. Recordings from a person’s point of view can capture how humans manipulate objects and perform everyday tasks at a much larger scale, although translating human video into actions that a specific robot can execute remains a technical challenge.
This creates a market in which the amount of data alone is not sufficient. Robot developers also need to know how data was recorded, what it contains, whether licensing permits its intended use and how closely it matches the robot and task they are developing.
Kinetic Blocks is attempting to position itself between the organisations collecting that data and the developers that need it. At this stage, however, the marketplace is still small and operating in beta. Its development will therefore also provide an indication of whether robotics training data can develop into a broader commercial market rather than continuing to be collected mainly through proprietary programmes and bilateral agreements.
Kinetic Blocks says it is expanding its supplier network and plans to build demand across Europe, the United States and Asia.
