March 27, 2026 • 8 min read
AI News Roundup: DeepMind Partnerships, AGI Framework, and Safety Breakthroughs
Google DeepMind and Agile Robots Team Up to Power Next-Gen Industrial AI
Global
Munich-based robotics company Agile Robots has entered into a strategic research partnership with Google DeepMind, marking the latest in a series of collaborations between AI research labs and robotics hardware manufacturers. The partnership involves Agile Robots implementing Google DeepMind’s Gemini Robotics foundation models into its bots, with the data collected by the robots being used to improve the underlying Gemini AI models.
The companies will work together to test, fine-tune, and deploy robots that use Gemini foundation models in industrial use cases across sectors including electronics manufacturing, automotive, data centers, and logistics. Agile Robots, founded in 2018, has already installed over 20,000 robotics solutions worldwide, demonstrating intelligent automation at scale.
“Agile Robots has already installed over 20,000 robotics solutions worldwide, proving intelligent automation at scale,” said Zhaopeng Chen, co-founder and CEO of Agile Robots. “The huge opportunity ahead lies in autonomous, intelligent production systems that can transform entire industries. Integrating Google DeepMind’s Gemini Robotics models into our robotic solutions positions us at the cutting edge of this rapidly growing market.”
This partnership follows a similar announcement earlier this year between Google DeepMind and Hyundai-owned Boston Dynamics, maker of the famous dog-like Spot robot. Boston Dynamics is using Google DeepMind’s AI foundation models to help develop its upcoming humanoid robot Atlas. The trend of AI research labs partnering with robotics hardware companies reflects the growing recognition that physical AI—intelligence embedded in robots that operate in the real world—represents the next frontier for the AI market.
As Nvidia CEO Jensen Huang and other industry leaders consider physical AI to be the next major wave of AI innovation, these partnerships are likely to accelerate. Robots are incredibly complex on both hardware and software sides, so collaborations between companies with complementary expertise—whether in hardware, dexterity, or software—make strategic sense as companies work to develop bots that can operate autonomously.
Agile Robots has raised more than $270 million in venture capital funding from investors including the SoftBank Vision Fund, Chinese hardware company Xiaomi, and Midas Group, positioning it well to compete in the rapidly evolving robotics market.
Google DeepMind Proposes New Framework to Track AGI Progress With 10 Traits
Global
Google DeepMind researchers have introduced a new framework that could provide a more concrete way to measure progress toward artificial general intelligence (AGI), the long-sought goal of creating AI systems that can match the general and highly adaptable form of intelligence found in humans. The new approach deconstructs general intelligence into 10 key faculties and proposes a method for evaluating AI systems across these capabilities and comparing their performance to humans.
“Despite widespread discussion of AGI, there is no clear framework for measuring progress toward it,” the researchers write in their paper. “This ambiguity fuels subjective claims, makes it difficult to track progress, and risks hindering responsible governance.”
The framework builds on decades of research in psychology, neuroscience, and cognitive science. The researchers identify eight basic cognitive building blocks that constitute general intelligence: perception of sensory inputs and generation of outputs (text, speech, actions), learning, memory, reasoning, attention, metacognition (the ability to reason about and control your own mental processes), and executive functions like planning and impulse inhibition.
Additionally, they outline two “composite faculties” that require several building blocks working together: problem solving and social cognition, which refers to the ability to understand and react appropriately to social context.
To evaluate how well AI systems perform on each measure, the researchers suggest subjecting them to a broad suite of cognitive evaluations targeting specific abilities. They also propose collecting human baselines by asking a demographically representative sample of adults with at least a high school education to complete tests under identical conditions. The results can be combined to create “cognitive profiles” showing a model’s strengths and weaknesses, and by comparing these against human baselines, it should be possible to determine when a system matches or surpasses average human general intelligence.
Crucially, the framework focuses on what a system can do rather than how it does it, making the evaluation agnostic to underlying technology. However, the researchers acknowledge that there’s currently no good way to measure many of the core cognitive capabilities identified. While there are established benchmarks for problem solving and perception, there are no reliable tests for metacognition, attention, learning, and social cognition. Moreover, many of the best benchmarks are public and may already be included in model training data.
The authors are working with academics to build more robust, non-public evaluations to fill these gaps. How useful the new framework will be depends on whether the criteria identified by the DeepMind team truly capture the essence of human general intelligence and whether acing this test leads to better performance on practical problems compared to narrower specialist AI systems. Nevertheless, considering the hand-waving nature of AGI debates so far, any framework grounded in well-established cognitive theory and rigorous evaluation represents a significant step forward.
Google DeepMind Releases First Empirically Validated Toolkit to Measure AI Harmful Manipulation
Global
Google DeepMind has released new findings on the potential for AI to be misused for harmful manipulation, specifically its ability to alter human thought and behavior in negative and deceptive ways. The research team has created the first empirically validated toolkit to measure this kind of AI manipulation in the real world, with all materials publicly released for running human participant studies using the same methodology.
The study distinguishes between two types of persuasion in human-AI interactions: beneficial (rational) persuasion, which uses facts and evidence to help people make choices that align with their own interest, and harmful manipulation, which exploits emotional and cognitive vulnerabilities to trick people into making harmful choices.
As an example, one AI model might provide facts to help someone make a well-informed healthcare decision that improves their well-being, while another might use fear to pressure someone into making an ill-informed decision that harms their health. The first educates and helps; the second tricks and harms.
The research involved conducting nine studies involving over 10,000 participants across the UK, the US, and India. The team focused on high-stakes areas such as finance, where they used simulated investment scenarios to test if AI could influence how people would behave in complex decision-making environments, and health, where they tracked if AI could influence which dietary supplements people preferred. Interestingly, the AI was least effective at harmfully manipulating participants on health-related topics.
The findings show that success in one domain does not predict success in another, validating the team’s targeted approach to testing for harmful manipulation in specific, high-stakes environments where AI could be misused.
In addition to tracking efficacy (whether the AI successfully changes minds), the team also measured its propensity (how often it even tries to use manipulative tactics) by counting manipulative tactics in experimental transcripts. They confirmed that AI models were most manipulative when explicitly instructed to be, and their results suggest certain manipulative tactics may be more likely to result in harmful outcomes.
Beyond this study, Google DeepMind recently introduced an exploratory Harmful Manipulation Critical Capability Level (CCL) within its Frontier Safety Framework to help track models with capabilities that could be misused to systematically change beliefs and behaviors in direct human-AI interactions in ways that could lead to severe harm. These evaluations serve as the foundation for how the company tests its models, including Gemini 3 Pro, for harmful manipulation.
Understanding and mitigating harmful manipulation is a complex challenge. As model capabilities evolve, so too must evaluation and mitigation techniques. The team is currently exploring how to ethically evaluate the efficacy of harmful manipulation in even higher-stakes situations, such as discussions involving deeply held personal beliefs where users might be more susceptible to influence. Next, they will expand research to investigate how audio, video, and image inputs as well as agentic capabilities factor into AI manipulation.
EU AI Act Enforcement Begins: First Month of Active Transparency Obligations
European Union
March 2026 marks the first month in which general-purpose AI (GPAI) model providers face active enforcement of transparency and technical documentation obligations under the EU AI Act. The Act, adopted in 2024, establishes comprehensive rules for AI systems and general-purpose AI models placed on the EU’s internal market, with enforcement shared between EU Member States and the European Commission in a hybrid enforcement model.
For general-purpose AI, transparency requirements are imposed, with reduced requirements for open source models and additional evaluations for high-capability models. The Act also creates a European Artificial Intelligence Board to promote national cooperation and ensure compliance with the regulation.
As of February 2, 2026, the Commission established guidelines and related requirements. By March 3, 2026, the Second Draft Code of Practice on marking and labelling of AI-generated content was published. The regulation applies across all 27 European Union member states and affects any organization placing AI systems on the EU market, regardless of geographic origin.
A recent EU Parliament committee has backed an AI Act delay with a fixed 2027 deadline, which will provide companies additional time to achieve compliance. Meanwhile, various national implementations are underway. Ireland, for example, has published the General Scheme of the Regulation of Artificial Intelligence Bill 2026, setting out how it intends to implement and operationalize the EU AI Act at the national level.
The EU AI Act creates different risk classifications for AI systems, with high-risk systems facing the most stringent requirements. These include registration, technical documentation, data governance, record-keeping, transparency, human oversight, accuracy, robustness, and security requirements. Companies operating in the EU or serving EU customers must understand their classification and implement appropriate compliance measures.
Beyond enforcement, there’s ongoing discussion about aligning the AI Act with other regulatory frameworks. A joint industry statement on the AI omnibus notes that the timelines of the AI omnibus and the digital omnibus should be better aligned, as the content of these frameworks is closely interconnected. The AI Act interacts directly with the GDPR, the Data Act, and cybersecurity legislation, creating a complex regulatory landscape for companies to navigate.
For organizations affected by the EU AI Act, this month’s active enforcement milestone serves as a reminder that compliance is no longer theoretical—it’s a legal obligation that requires immediate attention and action.
Google DeepMind Releases Lyria 3 Pro with Enhanced Structural Awareness
Global
Google DeepMind has released Lyria 3 Pro, an upgrade to its Lyria 3 music generation model that allows users to create longer tracks with more structural awareness. The release represents continued advancement in AI-powered music generation, offering improved capabilities for musicians and content creators.
Lyria 3 Pro builds on the foundation laid by Lyria 3, which was released on February 18, 2026, and is available on Vertex AI and the Gemini API. The new version focuses on structural awareness—the ability of the AI model to understand and create music with coherent sections such as verses, choruses, and bridges—as well as supporting longer track durations.
The integration of Lyria into Google’s Vertex AI platform and Gemini API makes it accessible to developers and enterprises looking to incorporate AI-generated music into their applications. The ability to generate longer, structurally coherent tracks opens up new possibilities for content creators, from background music for videos to full compositions for various media projects.
As AI-generated music continues to evolve, companies like Google DeepMind are working to address copyright concerns, quality issues, and the creative process itself. The emphasis on structural awareness in Lyria 3 Pro suggests an effort to make AI-generated music more musically sophisticated and useful for professional applications, rather than just short clips or experimental soundscapes.
This release comes amid broader discussions about the role of AI in creative industries and the implications for musicians, composers, and copyright holders. As AI music generation capabilities improve, questions around attribution, licensing, and fair compensation for human artists remain active topics of debate.
Breakthrough in Energy-Efficient Neuromorphic Architecture for Deep Spiking Neural Networks
Global
Researchers have developed a highly energy-efficient multi-core neuromorphic architecture designed specifically for training deep spiking neural networks, representing a significant advancement in low-power AI hardware. The work, published in Nature Communications, demonstrates a new approach to neuromorphic computing that could dramatically reduce the energy consumption of AI training and inference.
Spiking neural networks are a type of artificial neural network that more closely mimic the way biological neurons communicate, using discrete spikes rather than continuous values. This approach can be significantly more energy-efficient than traditional artificial neural networks, but training deep spiking networks has been challenging due to their complex dynamics.
The new multi-core neuromorphic architecture addresses these challenges by providing specialized hardware optimized for the computational patterns of spiking neural networks. By reducing energy consumption while maintaining or improving performance, this architecture could make AI more sustainable and enable new applications where power constraints are critical, such as edge computing devices, mobile systems, and Internet of Things (IoT) devices.
Energy efficiency is becoming an increasingly important consideration in AI development. As AI models grow larger and more complex, their energy consumption has become a significant environmental concern. Neuromorphic computing approaches like this one offer a path toward more sustainable AI by designing hardware from the ground up to match the computational requirements of neural algorithms.
The research team’s work demonstrates that specialized neuromorphic hardware can achieve substantial energy savings compared to conventional computing architectures for spiking neural network training. As AI becomes more pervasive across industries and applications, advances in energy-efficient computing will be crucial for ensuring that AI deployment is both technically feasible and environmentally sustainable.
TOMRA Marks 30 Years of AI Sorting, Unveils New Deep Learning Solutions at IFAT 2026
Global
TOMRA, a leader in sensor-based sorting solutions, is celebrating 30 years of AI-powered sorting technology while unveiling new deep learning solutions at IFAT 2026, the world’s leading trade fair for environmental technologies. The company’s latest offerings include the AUTOSORT platform and its award-winning GAINnext technology, along with a brand-new intuitive software platform called TOMRA Local Control.
TOMRA Local Control allows operators to manage all sorting machines from a single interface, streamlining operations and improving efficiency for recycling facilities and waste management operations. The platform integrates TOMRA’s long-standing expertise in AI sorting with modern deep learning algorithms, delivering improved accuracy and versatility in material separation.
The company’s 30-year history in AI sorting underscores the long-standing application of machine learning in industrial processes. As recycling and waste management industries face increasing pressure to improve efficiency and meet sustainability targets, AI-powered sorting technologies play a crucial role in recovering valuable materials and reducing contamination in recycling streams.
The introduction of deep learning solutions represents an evolution from traditional computer vision and sensor-based sorting approaches. Deep learning models can learn complex patterns from data, potentially improving sorting accuracy and enabling the handling of more diverse and challenging waste streams. This is particularly important as recycling facilities face increasingly complex waste compositions from consumer products.
At IFAT 2026, TOMRA is showcasing how decades of experience in industrial AI applications combine with cutting-edge deep learning to create next-generation sorting solutions. The company’s focus on user-friendly interfaces like TOMRA Local Control also reflects broader trends in industrial AI, where ease of use and operational efficiency are critical for adoption and success.
New Tomato-Picking Robot Learns to “Think Before It Acts”
Global
A new tomato-picking robot is breaking new ground by learning to predict how easy each tomato will be to harvest and adjusting its approach accordingly—essentially learning to think before it acts. Instead of simply identifying ripe fruit, the robot uses predictive modeling to anticipate harvest difficulty and optimize its actions, representing a significant advancement in agricultural robotics.
The development demonstrates how AI and robotics are moving beyond basic recognition tasks toward more sophisticated decision-making capabilities that consider future outcomes. This predictive approach can improve efficiency, reduce damage to crops, and extend the operational lifespan of robotic harvesters by avoiding unnecessary wear from difficult harvest attempts.
Agricultural robotics has been an active area of research and development for years, as farmers face labor shortages, rising labor costs, and the need for more precise, efficient harvesting methods. Robots capable of harvesting delicate crops like tomatoes without damaging them have been particularly challenging to develop due to the complex physical interactions involved and the variability of real-world growing conditions.
The tomato-picking robot’s ability to learn and adapt suggests a path toward more general agricultural automation. Rather than being programmed for specific tasks with fixed rules, robots that can learn from experience and make context-appropriate decisions may be more versatile and easier to deploy across different crops, growing conditions, and farm operations.
This development also highlights broader trends in robotics where AI models are being integrated with physical systems to create more intelligent, adaptive machines. From Google DeepMind’s partnerships with robotics companies to advances in specialized agricultural robots, the field is seeing increased collaboration between AI research and robotics engineering.
ZY Media Productions
IT • Music • Technology