SICS AI's EIC Accelerator Promises AGI for Robots by 2026

An industrial robot is excellent at repeating a motion and helpless when the world changes. That is the gap SICS AI, based in Stockholm with an office in Palo Alto, says it is closing with a foundation model that can control any robot regardless of its geometry or kinematics, learn online from each interaction, and build what the company calls a world map at the centre of the model. The CORDIS summary lists the four limitations it targets: robustness, a structured world model, online interaction, and multi-step reasoning and planning.

The EIC Accelerator selected the project, HYPER, at its October 2024 cut-off for blended finance, and CORDIS records a grant of €2,389,165.64 from July 2025 to December 2026, an 18-month project and the shortest of this cut-off. The company's own website is less hedged than its grant: "We will deliver AGI in 2026."

The claim, stated plainly

SICS AI's position is that today's AI is the wrong shape, not too small: "Contemporary AI operates through statistical pattern recognition instead of actual reasoning, inherently limiting its ability to reach human-level intelligence. No amount of computational power or data can bypass this limitation. We have uncovered the key missing link in AGI development." The argument it offers for its approach is biological: an intelligent organism navigates, navigation requires planning and reasoning, so a world map belongs at the centre of the model, which then learns like a child.

Nothing in public supports or refutes any of this. There is no paper, no benchmark result, no demonstration video with measured performance, and Swedish coverage notes that the HYPER architecture's details are being withheld pending patents. What exists is a product name, Robot Brain One, for warehouse pick-and-place robots, and a Swedish press report of a robot picking goods at an industrial site outside Helsingborg.

The LUMI allocation

The one externally verifiable asset is compute. SICS AI has an allocation on LUMI, the EuroHPC supercomputer in Finland, to accelerate development of its robot control model, according to the Swedish national competence centre for high-performance computing. For a small company training a foundation model, access to a top-tier European machine is a material advantage and a form of validation, since allocations are awarded on technical review.

The EIC grant has its own logic in the project text, which is unusually frank about what the money is for beyond engineering: EIC support should lower the barrier for venture capital investors to co-invest alongside the EIC, and the grant can serve as independent quality approval and due diligence for investors. In its own words, the project expects the grant to act as that signal for investors.

1. Access to training computeGrantedLUMI allocation, reported September 2025.
2. A robot running the modelReportedSwedish press describes a picking robot near Helsingborg; no performance data.
3. Published evidence of the architectureNoneDetails withheld pending patents; no benchmark or paper.
4. AGI in 2026Not credible on current evidenceNo public result of any kind supports it.

The round that is supposed to follow

Realtid reports that SICS AI has raised more than SEK 58 million and that the European Investment Bank has pledged to take part in a planned round of SEK 110 million. Neither the company nor the EIB has announced that round as closed, and no EIC Fund equity has been announced either. Third-party databases give conflicting totals for the company's funding, so none is used here. At about €2.39 million (roughly SEK 26 million), the EIC grant is less than half the equity the company is reported to have raised so far.

Sovereign AI, with the emphasis on sovereign

The company's press page shows what it is selling to policymakers as much as to warehouses: articles on Sweden bidding for AI leadership, on sovereign computing capacity, and on why industry needs a world-model AI rather than a language model. That is a real European argument, since robotics foundation models are being built mostly in the US and China, and an allocation on a European supercomputer is exactly the sort of asset the EU wants used. It also means the company's central technical claim remains unpublished while its public messaging addresses policymakers.