Imagine a robot that learns like a curious toddler — it plays around, discovers what it can do, and remembers those skills for later. Right now, every time you want a robot to do something new, an engineer has to program it from scratch. The GOAL-Robots team built robots that set their own practice goals and teach themselves a growing library of skills — so when you finally ask them to tidy up a messy table, they already know how to grab, move, and place objects because they figured it out on their own. It's the difference between a robot that follows a script and one that genuinely adapts.
Robots in warehouses and factories today need to be painstakingly programmed for every new task, product, or layout change. When product mixes shift or environments change unexpectedly, companies face costly reprogramming cycles and downtime. There is no mainstream solution for robots that can genuinely teach themselves new skills and adapt to unforeseen situations without constant human re-engineering.
The team built 4 robot demonstrators (GOAL1 through GOAL4) showing progressively more autonomous learning. The final GOAL4 demonstrator solved a real-life tidy-up challenge using only skills the robot had taught itself, without knowing the specific objects or configurations in advance. Across 35 deliverables, the project produced computational architectures and algorithms for goal self-generation and open-ended skill acquisition.
If you are a warehouse operator dealing with constantly changing product mixes and layouts — this project developed robots that autonomously learn to handle, sort, and place objects without being reprogrammed for each new SKU. The GOAL4 demonstrator showed a robot solving a real-life tidy-up challenge by relying on skills it taught itself, without knowing the specific objects or configurations in advance.
If you are a manufacturer dealing with frequent product changeovers that require costly robot reprogramming — this project developed learning architectures where robots build up skill libraries through autonomous practice. Across 4 demonstrators, the team showed robots acquiring increasingly complex manipulation abilities with little human supervision, potentially cutting changeover downtime.
If you are a robotics company struggling to make your robots handle the unpredictability of real homes or offices — this project produced computational architectures that let robots self-generate goals and learn flexible skills for conditions unforeseeable at design time. The 35 deliverables include algorithms and architectures that could be integrated into next-generation service robots.
No pricing or licensing terms are publicly available. The project was led by Consiglio Nazionale delle Ricerche (Italian National Research Council) with 7 academic and research partners. Any licensing would need to be negotiated directly with the consortium, likely through CNR as coordinator.
The 4 demonstrators (GOAL1 through GOAL4) showed progressively more complex autonomous learning, culminating in a robot solving a real-life tidy-up challenge. However, all demonstrations were conducted in research settings with no industrial partners involved. Scaling to factory or warehouse environments would require significant additional engineering.
IP is held by the 7 consortium partners across 3 countries (Italy, Germany, France), primarily universities and research organizations. Under Horizon 2020 RIA rules, each partner owns the IP they generated. Access for commercial use would require a licensing agreement.
The core innovation is minimal human intervention — the robot self-generates its own practice goals using intrinsic motivation mechanisms. The GOAL3 demonstrator specifically showed robots learning several different skills with little supervision across several different conditions. However, initial hardware setup and environment preparation are still required.
The project ran from November 2016 to April 2021 and is now closed. All 35 deliverables and 4 demonstrators were completed. The algorithms and architectures developed are available for further development or licensing, but no commercial product has been announced.
Based on available project data, regulatory compliance for commercial deployment was not within scope. The project focused on fundamental learning algorithms rather than industrial safety certification. Any company adopting this technology would need to handle CE marking and machinery directive compliance separately.
The consortium of 7 partners across Italy, Germany, and France is entirely academic and research-focused: 4 universities, 2 research organizations, and 1 other entity, with zero industrial partners and zero SMEs. This is typical of FET Open projects that push fundamental science rather than near-market development. The coordinator, Italy's National Research Council (CNR), is one of Europe's largest public research bodies. For a business buyer, this means the technology is scientifically strong but has not been tested with real industrial requirements, integration constraints, or commercial viability in mind. Any adoption path would need a technology transfer or joint development partner to bridge the gap from lab to production.
Consiglio Nazionale delle Ricerche (CNR), Italy — reach out to the robotics or AI division for licensing inquiries
Want to explore whether GOAL-Robots' self-learning technology fits your automation challenges? SciTransfer can arrange a direct introduction to the research team and help assess commercial fit.
