Technical Challenges in Brain-Inspired Robotics

Drawing insights from biological signal processing, neuromorphic computing promises a substantially lower power solution to improve energy efficiency of visual odometry (VO) in robotics. Published in Nature Machine Intelligence, this novel approach develops a VO algorithm built from neuromorphic building blocks called resonator networks. Demonstrated on Intel’s Loihi neuromorphic chip, the network generates and stores a working memory of the visual environment, while at the same time estimating the changing location and orientation of the camera. The system outperforms deep learning approaches on standard VO benchmarks in both precision and efficiency – relying on less than 100,000 neurons without any training. This work is a key step in using neuromorphic computing hardware for fast and power-efficient VO and the related task of simultaneous localization and mapping (SLAM), enabling robots to navigate reliably.   A companion paper explores how the neuromorphic resonator network can be applied to visual scene understanding. By formulating the generative model based on vector symbolic architectures (VSA), a scene can be described as a sum of vector products, which can then be efficiently factorized by a resonator network to infer objects and their poses. The work demonstrates a new path for solving problems of perception and many other complex inference problems using energy efficient neuromorphic algorithms and Intel hardware. Congratulations to researchers from the Institute of Neuroinformatics, University of Zurich and ETH Zurich, Accenture Labs, Redwood Center for Theoretical Neuroscience at UC Berkeley, and Intel Labs.   Learn more about neuromorphic VO: https://lnkd.in/gJCVVMCz   Learn how the VSA framework was developed for neuromorphic visual scene understanding based on a generative model (companion paper): https://lnkd.in/gjAENfpp   #iamintel #Neuromorphic #Robotics

Humanoid Robots Enter the Next High-Stakes Phase of the AI Economy Bloomberg Businessweek explores the growing surge of investment and attention surrounding humanoid robotics as technology companies attempt to position physical AI systems as the next major frontier after generative artificial intelligence. While demonstrations of robots dancing, running, and performing simple tasks have generated enormous public excitement, the article argues that creating economically viable humanoid workers remains far more difficult than current hype suggests. The article centers on Persona AI, a Houston-based startup led by robotics veteran Nicolaus Radford, which is developing humanoid robots intended for industrial deployment in environments such as shipyards, energy infrastructure, and construction sites. The company’s long-term vision is to create a robotic labor platform capable of renting humanoid workers to employers worldwide, effectively establishing a robotic staffing industry. A major challenge highlighted in the article is the extraordinary complexity of replicating human dexterity, especially the capabilities of the human hand. Radford describes the human hand as “hypercompetitive,” capable of handling delicate precision tasks and heavy industrial actions with remarkable sensory feedback and adaptability. Replicating that versatility in robotics requires breakthroughs in mechanics, sensors, actuators, energy systems, software coordination, and real-world AI reasoning. The article also emphasizes that current humanoid robots remain expensive, power-intensive, and operationally limited outside carefully controlled environments. Viral demonstrations often mask the enormous engineering barriers associated with balance, object manipulation, autonomous navigation, safety, durability, and long-duration deployment in unpredictable real-world conditions. Many systems still rely heavily on teleoperation, scripted routines, or tightly constrained task environments. At the same time, investor enthusiasm continues accelerating as companies market humanoid robotics as a future solution for labor shortages, dangerous industrial work, aging populations, logistics automation, and military or disaster-response applications. Advances in AI models, machine vision, and edge computing are improving robot adaptability, but translating laboratory capabilities into scalable commercial deployment remains a massive undertaking. Key takeaways from the article suggest humanoid robotics may eventually become a transformational industry, but the sector is still early in its maturity cycle. The current excitement reflects both genuine technological progress and speculative expectations fueled by the broader AI boom. Success will likely depend not on viral demonstrations alone, but on whether companies can deliver reliable, economically scalable systems capable of operating safely and productively in complex human environments over long periods of time.

AGI Hype is Muting as Reality Imposes Itself Since 1956, AGI has been “five years away” according to its promoters, but now even the loudest voices are quieting down as skeptical voices of technologists and economists highlight the failures, such as the Anthropic disaster when an AI agent was put in charge of a store, producing technical and financial collapse. Promotors focused on computational models of the brain and have ignored the growing biological evidence that their perspective is myopic. Last year neuroscientists at Google imaged 1cmm of the brain, finding 57k cells and 150M synapses, 1.4 petabytes of data for one millionth of the brain. The combinatory complexity is mind-boggling, but that is not the complete story. In recent years, the study of astrocytes, the most abundant and diverse glial cells in the central nervous system, has revealed that their behavior includes chemical signals that modulate the activities of neurons, scaling the brain interaction complexity as well as the number of our unanswered questions. In 2025 a paper titled "Intercellular communication in the brain through a dendritic nanotubular network" (DOI: 10.1126/science.adr7403) has added other dimensions of complexity. From editor’s summary Synaptic connections mediate classical intercellular communication in the brain. However, recent data have demonstrated the existence of noncanonical routes of interneuronal communication mediating the transport of materials including calcium, mitochondria, and pathogenic proteins such as amyloid beta (Aβ). Using super-resolution and electron microscopy, Chang et al. identified and characterized structures called nanotubular bridges that connect dendrites in the brain. It should be clear that intelligence cannot be replicated by simply modeling neuronal inputs and outputs and scaling the model using statistical algorithms. But the issue is not limited to the scale and complexity of the architecture. It has been known for some time that the brain contains networks of stimulation and suppression cells that leverage memory and the brain’s world model to compute the relevance of inputs and filter what is contextually irrelevant, as well as categorizing a single new data point. In recent years concept cell networks have been identified as the primary method used by the brain to model the real world. Concept cells represent ideas without details and help explain how infants learn as many languages as they are exposed to regularly. World models are critical to developing computed intelligence. Yann LeCun and Google DeepMind are focusing research on world models. But the complexity of the real world including the complexity of the human brain underscore the challenge. We comprehensively cannot model reality, but we can apply principles discerned from biology to model problem domains.

🎄What a way to finish this year! 📝Our paper 'A Brain-Controlled Quadruped Robot: A Proof-of-Concept Demonstration' is just got published! MIT Media Lab Massachusetts Institute of Technology If you want to learn more on how to build: 🤖 autonomous 📱mobile 👁️eyes-free system using 🧠brain-sensing glasses and Boston Dynamics Spot, check our paper: https://lnkd.in/e-zjbMmK We call this system Ddog 🐶 👓 🧠 The goal of Ddog was to design an autonomous application that would enable a user to control a Spot robot based on the user’s input via brain only and give feedback to the user and their caregiver using voice via the app 📱 Ddog is designed to work either fully offline, or fully online. The online version has a more advanced set of machine learning (ML) models, as well as better fine-tuned models. Our paper also features an overview of the state of the art in 🧠 sensing and 🤖as of 2023. We explain the challenges of portability, comfort, UI/UX for the users who would wear brain computer-enabled devices, with our proposed solution (I) being wearable form-factor with a setup time < 2 min. Video trailer of Ddog is available here: https://lnkd.in/ey44XJGq AttentivU measures the user’s electroencephalography (EEG - 🧠 ) and electrooculography (EOG - 👁️) activity of the user from the electrodes embedded in the glasses’ frame. The user mentally responds to a series of questions with YES/NO answers, using their brain only. Each question–answer pair has a pre-configured set of actions for Spot. For instance, Spot was prompted to walk across a room, pick up an object, and retrieve it for the user (i.e., bring a bottle of water) when a sequence resolved to a YES response with a success rate of 83.4%. To the best of our knowledge, this is the first integration of wireless, non-visual-based BCI system with Spot in the context of personal assistant use cases. While Ddog is an early prototype, future iterations may embody friendly and intuitive cues similar to service dogs 🐕🦺 [IN-DEPTH VIDEO OVERVIEW] Ddog: The World’s First Brain-Controlled Spot Robot is available here: https://lnkd.in/eP22akRp Thank you so much to my amazing co-authors 🔬Eugene Hauptmann and Yasmeen Hmaidan - this work would not be possible without them! And of course, thank you to Spot!

New AI can ID brain patterns related to specific behavior As you are reading this story, your brain is involved in multiple behaviors. Perhaps you are moving your arm to grab a cup of coffee, while reading the article out loud for your colleague, and feeling a bit hungry. All these different behaviors, such as arm movements, speech and different internal states such as hunger, are simultaneously encoded in your brain. This simultaneous encoding gives rise to very complex and mixed-up patterns in the brain’s electrical activity. Thus, a major challenge is to dissociate those brain patterns that encode a particular behavior, such as arm movement, from all other brain patterns. For example, this dissociation is key for developing brain-computer interfaces that aim to restore movement in paralyzed patients. When thinking about making a movement, these patients cannot communicate their thoughts to their muscles. To restore function in these patients, brain-computer interfaces decode the planned movement directly from their brain activity and translate that to moving an external device, such as a robotic arm or computer cursor. The researchers developed a new AI algorithm that addresses this challenge. The algorithm is named DPAD, for “Dissociative Prioritized Analysis of Dynamics.” “Our AI algorithm, named DPAD, dissociates those brain patterns that encode a particular behavior of interest such as arm movement from all the other brain patterns that are happening at the same time,” the author said. “This allows us to decode movements from brain activity more accurately than prior methods, which can enhance brain-computer interfaces. Further, our method can also discover new patterns in the brain that may otherwise be missed.” #ScienceMission #sciencenewshighlights https://lnkd.in/gDx6jrAV

What will it take to go from LLMs to truly intelligent agents? Large Language Models (LLMs) have become the engines of AI, but the vehicles we build on top of them—intelligent agents—still have a long road ahead. I recently explored a deep and thought-provoking paper, Advances and Challenges in Foundation Agents (Bang Liu et al., 2025), which outlines a modular, brain-inspired architecture for agents and calls attention to four major pillars of progress:- 1. Modular Cognition:- Drawing from neuroscience, agents need integrated systems—memory, world models, emotion, and reward mechanisms—to operate autonomously in dynamic environments. 2. Self-Evolution:- The next frontier is continual learning and LLM-driven self-optimization, allowing agents to grow their capabilities without constant human intervention. 3. Multi-Agent Collaboration:- Just like humans, agents must interact, learn, and evolve as teams, adapting communication and workflow protocols to collaborate effectively. 4. Safety & Ethics:- Perhaps the most critical challenge—how do we build secure, aligned, and trustworthy agents that avoid hallucinations, prompt injections, or misaligned behaviors? The authors liken LLMs to jet engines, and agents to the aircraft we build with them. But today’s aircraft still struggle with takeoff—long-term memory, emotional grounding, and goal-directed action remain underdeveloped. One key insight I loved:- The human brain—with its modular regions like the frontal lobe (planning), hippocampus (memory), and cerebellum (motor learning)—offers a template for designing better agents. But we need to go beyond imitation, toward augmentation. If you’re building in the space of AI agents or just curious about how LLMs might evolve into autonomous collaborators, this is a must-read. 📃Full paper here:- https://lnkd.in/dMW2QdM2 💻GitHub:- https://lnkd.in/d9Wph8_P Would love to hear your take—what do you think is the most underexplored function in AI agents today? #AI #FoundationAgents #LLMs #AgenticAI #ArtificialIntelligence #NeuroscienceInspired #ResponsibleAI #CollectiveIntelligence

Brain Size Doesn't Predict Intelligence, Movement Does Current AI systems require massive computational resources to achieve basic pattern recognition tasks that insects accomplish with brains smaller than sesame seeds. University of Sheffield researchers discovered that bees use flight movements and body wiggles to help their brains learn and recognize visual patterns with remarkable accuracy, achieving 96-98% accuracy versus 60% for stationary observation. Organizations developing AI systems have focused on scaling up neural networks rather than optimizing how information gets processed. Bee-inspired computational models demonstrate that intelligence emerges from dynamic interaction between brains, bodies, and environments rather than pure processing power, suggesting a fundamentally different approach to building efficient AI. The computational model shows that just 16 lobula neurons prove sufficient for complex pattern discrimination tasks, while bees scanning only the lower half of patterns dramatically outperform stationary observation methods. This reveals that active perception through movement creates sparse, decorrelated neural responses where specific neurons activate only for particular visual features. Robotics applications could benefit enormously from this bio-inspired approach, enabling systems that actively shape their sensory input through movement rather than passively processing massive datasets. Future robots can become smarter and more efficient by using movement to gather information, rather than relying on massive computing power. Implementation challenges include developing movement strategies that optimize information gathering while maintaining energy efficiency. The research suggests that intelligence requires embodied interaction with environments, not just larger neural networks or faster processors. Bio-inspired AI represents a pathway toward more sustainable artificial intelligence that achieves better performance with dramatically reduced computational requirements, fundamentally changing how we approach machine learning architecture design. 🔗https://lnkd.in/ePcWpYK5

CL1 'biological computer' unit by Cortical Labs. Inside this device, living human neurons are cultured on a microchip and kept alive with built-in life support systems. Recently, I came across news about Cortical Labs, an Australian startup has created the first commercially available biological computer, intriguingly called "CL1." But what exactly is a biological computer, and why should we pay attention? Unlike the silicon-based computers we have all become familiar with - built upon chips and code - this biological computer incorporates living human neurons into its hardware. Essentially, they have cultivated real human brain cells onto a specialized silicon surface, allowing these cells to naturally form networks that communicate via electrical signals. These signals are then captured and relayed back and forth between biological and electronic components, enabling real-time interactions between biology and technology. What's particularly interesting is how differently this "brain-on-a-chip" approaches problem-solving compared to standard artificial intelligence models. Traditional AI relies on extensive training data, heavy computational power, and explicit programming. In contrast, biological computers like CL1 learn more organically, adapting quickly to new situations using far less energy. Because neurons naturally reorganize themselves, they can respond and adapt in real time - something current AI technologies struggle with. This innovation isn't just academic; it has practical implications across various fields. Medical researchers could harness biological computing to model diseases more accurately, reducing the reliance on animal testing and potentially accelerating drug discovery. Roboticists might use these adaptive networks to create robots capable of genuine, continuous learning and environmental adaptability - critical traits for tasks in unpredictable, real-world scenarios. Yet, exciting as this innovation is, it also raises significant ethical and practical questions. For instance, how do we responsibly manage technologies that integrate living human tissue with machinery? While the current systems are too simple for anything approaching consciousness, the very existence of such technology pushes us to clearly define boundaries and ethical guidelines early on. From a technical standpoint, scaling these biological computers for larger, more complex tasks also poses a considerable challenge. Keeping neurons alive and functioning reliably over extended periods, and integrating these "wetware" components effectively with conventional hardware, demands innovative engineering solutions. What do you think about this emerging blend of biology and technology? #innovation #technology #future #management #startups