AMI Scientist

10 European startups shaping our robotic future. August 18, 2026 Europe's robotics and physical AI ecosystem is expanding beyond traditional industrial automation to include humanoid robots, autonomous defence and security systems, software-defined manufacturing, and AI models designed to understand and act in the physical world. This list highlights some of Europe's most promising robotics startups founded between 2024 and 2026. Together, they demonstrate the growing convergence of artificial intelligence, robotics, autonomy and advanced engineering. While not all of these companies are pure-play robotics startups, each is developing technology that enables machines, autonomous systems or industrial infrastructure to operate more intelligently in real-world environments. Several have also attracted significant early-stage investment, reflecting growing investor interest in the technologies underpinning the next generation of physical automation. Advanced Machine Intelligence. Founded in Paris in 2026, Advanced Machine Intelligence develops AI systems that build abstract representations of physical and digital environments from sensor data, allowing intelligent systems to reason, predict outcomes and plan actions. Its work has potential applications across robotics, industrial automation, wearables and healthcare. Advanced Machine Intelligence raised approximately €890 million at a €3 billion pre-money valuation in one of Europe's largest-ever Seed rounds in March this year, Cambridge Aerospace. Operating from Cambridge in the United Kingdom, Cambridge Aerospace was founded in 2024 and develops air-defence technology for detecting and intercepting aerial threats. Its work includes drone and missile interceptors, radar systems and rocket propulsion technology, alongside manufacturing capabilities for its defence systems. Cambridge Aerospace has raised more than $630 million to date, which includes a Series C in August of this year at a €2.94 billion valuation. Flexion Robotics. Zurich is home to Flexion Robotics, a robotics company founded in 2024. It is developing an AI-powered autonomy software stack for humanoid robots, combining reinforcement learning, language-based reasoning, vision-language-action models and whole-body control. The company trains its systems in high-fidelity simulations to help robots learn complex tasks while reducing reliance on human demonstrations. Flexion Robotics has raised approximately €49.6 million to date. Generative Bionics. Founded in 2025 in Genoa, Italy, Generative Bionics develops humanoid robots and Physical AI systems for industrial and service environments. Its approach brings together robotics, artificial intelligence and engineering to support safe and reliable operation in real-world settings. Generative Bionics has raised €70 million in funding so far. Harmattan AI. Also emerging from Paris' growing defence technology ecosystem, Harmattan AI was founded in 2024 and develops AI-powered autonomous systems for military applications. The company's technology covers unmanned vehicles, sensor integration, mission planning and real-time decision-making, with systems designed for operations across land, air and maritime environments. Harmattan AI has raised approximately €209.2 million in funding including a Series B in January of this year. Hive Robots. In Kongens Lyngby, Denmark, Hive Robots is building humanoid robot workforce solutions for businesses. Founded in 2025, the company offers proof-of-concept programmes, robot-as-a-service models, software development and vertically integrated deployments. Hive targets sectors including logistics, retail, hospitality and healthcare across Scandinavia and Northern Europe. Humanoid. With its headquarters in London, Humanoid is developing general-purpose humanoid robots for industrial and commercial environments. Its technology combines proprietary humanoid hardware with its KinetIQ AI and software stack, initially targeting applications across manufacturing, logistics and warehousing. Humanoid raised €133 million in Series A funding in July 2026, reaching a €1.1 billion post-money valuation. Isembard. London-headquartered Isembard is building software-powered factories for high-precision manufacturing across aerospace, defence and energy. Founded in 2024, the company operates modular manufacturing units equipped with machine tools and robotics, producing precision components from materials including metals, plastics and composites. In March of this year, the company raised €43 million ($50 million) in Series A funding, less than 12 months after its Seed round. Rollo Robotics. Founded in Tallinn in 2025, Rollo Robotics develops autonomous monowheel security robots that use gyroscopic stabilisation technology. The company combines robotics, sensing and automation to support continuous security and monitoring applications. Rollo Robotics has so far raised €3.7 million in pre-Seed funding. Stateful Robotics. Emerging from Oxford's robotics research ecosystem, Stateful Robotics was founded in 2025 and develops a hardware-agnostic mission-control platform for mobile robots operating in dynamic environments. Its technology has been used with quadruped robots carrying out hazardous inspections, mobile service robots working in security and care settings, and underwater robots collecting environmental data. Stateful Robotics has raised €4.1 million. Summary. Taken together, these companies illustrate how Europe's robotics sector is developing beyond traditional industrial automation. Their approaches span foundational Physical AI, humanoid platforms, autonomous defence systems, security robots and software for coordinating robotic fleets, with applications across manufacturing, logistics, defence, healthcare and other physical environments. Funding levels vary considerably, but the scale of recent rounds is notable. Advanced Machine Intelligence secured close to €890 million in 2026, while Cambridge Aerospace raised more than €430 million across two rounds during the year. Harmattan AI's $200 million Series B and Humanoid's €133 million Series A further point to substantial capital moving into the intersection of AI, robotics and autonomous systems.

AdaJEPA - Yann LeCun's Lab builds a World Model that keeps learning while it acts. A world model that stops learning the moment it is deployed will plan against the wrong imagined future the moment reality drifts. AdaJEPA - new work from NYU's Agentic Learning AI Lab and AMI Labs by Ying Wang, Oumayma Bounou, Yann LeCun and Mengye Ren - makes the world model adapt inside the planning loop itself: at every model-predictive-control step the agent plans, acts, observes the resulting transition, and uses that transition as a self-supervised training signal before the next replan. The adaptation is deliberately lightweight - one gradient step per replan, a five-transition replay buffer, updates restricted to the final layers - yet it consistently improves planning under distribution shift: unseen object shapes, corrupted observations, changed physics and held-out maze layouts. In low-data regimes it can more than double a frozen model's success rate and even beat frozen models trained on far more data. Vivax unpack the plan-act-adapt-replan loop, the results, and why a clinical world model - deployed into hospitals where the data distribution never stops moving - should be built to keep learning too. What NYU released. NYU's Agentic Learning AI Lab and AMI Labs have released AdaJEPA: An Adaptive Latent World Model - new work by Ying Wang, Oumayma Bounou, Yann LeCun and Mengye Ren, with the paper on arXiv and code on GitHub. The one-line idea: instead of freezing a world model after training and hoping it generalizes, AdaJEPA adapts the model during deployment, inside the planning loop itself. Every action the agent executes produces an observed transition - what it saw, what it did, what happened next - and AdaJEPA uses that transition as a self-supervised training signal before the next replan. Plan, act, adapt, replan. The frozen world model problem. The problem it attacks is one every world-model system inherits. Latent world models make planning from high-dimensional observations tractable by predicting future states in a compact representation space - but they are usually kept frozen after training. When the deployed environment drifts from the training distribution - different object shapes, corrupted camera input, changed physics - the model's imagined futures go subtly wrong, and model-predictive control then optimizes actions for the wrong imagined future. The plan is coherent; the world it was planned for no longer exists. How AdaJEPA works. AdaJEPA starts from a pretrained JEPA world model - a sensory encoder, an action encoder, and a latent predictor - and plans with model-predictive control in latent space: roll the predictor forward over candidate action sequences and pick the one that lands closest to the goal's representation. The new part comes after each step. The observed transition goes into a small online buffer, and the model takes a gradient step on the latent prediction error before the next replan. The recipe is deliberately minimal: one gradient step per replan, a replay buffer of just five recent transitions, and updates restricted to the final layers of the visual encoder and predictor. This is test-time adaptation woven directly into the control loop - cheap enough to run at every step. What the results show. The results track the promise. In-distribution, adaptation helps when the frozen model is suboptimal and does no harm when it is already near-optimal. Under distribution shift, the gains are consistent: on PushObj with held-out object shapes, on PushT with blurred, noisy, darkened or recolored observations, and on PointMaze with altered physics and unseen layouts, each observed transition recalibrates the model before the next replan, improving success rates and pulling trajectories closer to shortest paths. The data-scaling result is the sharpest: when offline training data is limited, test-time adaptation can more than double the frozen model's success on seen shapes - and even outperform frozen models trained on far more data. The takeaway for world models. The visualizations make the mechanism legible: decoded rollouts from the adapted model still reconstruct sensibly on unseen shapes and corrupted views while retaining training-domain structure, which suggests adaptation works by exploiting shared latent structure and recalibrating predictions rather than wandering off the learned manifold. The authors' takeaway is bigger than the benchmarks: world models should continue learning during deployment rather than remain frozen after training. Perception and planning get more resilient when the model treats every step it takes as evidence about the world it is actually in. Why this matters for clinical world models. For medicine, this is not a robotics curiosity - it is the core problem restated. A clinical world model is deployed into permanent distribution shift: new scanners and assays, new hospitals and populations, evolving protocols and drug availability. A model frozen at training time starts aging the day it ships. AdaJEPA is a blueprint for the alternative - grounded models that recalibrate from the transitions they actually observe, under tight, auditable update budgets (one step, five transitions, final layers only) rather than open-ended retraining. That restraint matters doubly in healthcare, where every update must remain safe and traceable. It also extends the JEPA lineage Vivax has been tracking - from the original architecture through V-JEPA 2, Neuro-JEPA and IQ-JEPA - one more step toward Yann LeCun's vision of models that learn how the world works, and keep learning it.