Beijing’s Robot Kindergarten

A new robot kindergarten has opened in Beijing’s Shijingshan district to explore a different approach to training robots: allowing them to learn continuously through physical interaction, touch, trial and error, rather than relying primarily on human demonstrations. The facility was jointly established by tactile-sensing company Tashan Technology and a team led by Turing Award winner and reinforcement-learning pioneer Richard Sutton. Robots repeatedly attempt movements, receive feedback from their environment and modify their behaviour based on the results, for example, learning to avoid an obstacle after colliding with it. Sutton described the facility as a controlled environment where robots can learn about their own bodies and the physical world, making mistakes an essential source of training data.

The initiative reflects a broader effort to develop robots capable of continuous, autonomous learning after deployment, rather than having their behaviour largely fixed by pre-training, simulation or teleoperation. Researchers argue that robots could eventually generate their own experience-based datasets and potentially share what individual machines learn across common models, helping them adapt to situations not covered during initial training. Significant challenges remain, particularly hardware durability and the complexity of learning sophisticated humanoid behaviours. For now, autonomous experiential learning is expected to complement rather than replace imitation learning and simulation, while potentially becoming increasingly important as robots enter factories, services and homes.

More information:

https://www.globaltimes.cn/page/202609/1369570.shtml

Brain Implant Controlled Across Continents

Researchers from KAIST and Yonsei University have demonstrated a wireless brain implant that can be controlled across continents. Their system, called RAPIDO, received commands sent from Chicago to a laboratory in Daejeon, South Korea with an average response time of about 109 milliseconds. Designed for experiments with freely moving animals, the implant combines two capabilities: precisely delivering drugs into specific brain regions and providing targeted light stimulation for optogenetic experiments. Its refillable cartridge and internet-based control could enable longer-term studies while reducing the need for researchers to physically approach or repeatedly handle experimental animals.

In experiments with rats, RAPIDO successfully delivered controlled doses of cocaine into the nucleus accumbens over several weeks, producing measurable dose-dependent behavioral effects. In a separate experiment, researchers used the implant's light stimulation to activate the RhoA signaling pathway, preventing rats from developing the same learned preference for a cocaine-associated environment. The researchers stress that this does not constitute a treatment for addiction; rather, the work demonstrates a platform for remotely manipulating specific neural processes and studying their effects on behavior. Significant challenges must be addressed before such technology could be considered for human clinical applications.

More information:

https://www.sciencealert.com/a-brain-implant-in-korea-was-just-controlled-from-the-us

LLM-Personalised Autonomous Driving

Researchers at Delft University of Technology have developed an approach that allows passengers to personalise how an autonomous vehicle drives using natural-language commands. Instead of allowing a large language model (LLM) to control the vehicle directly, the system uses the LLM to interpret requests or indications that the passenger feels dizzy. These preferences are translated into adjustments to parameters used by a conventional, safety-aware motion planner, affecting characteristics such as speed, acceleration, steering smoothness and collision avoidance. The system also keeps the passenger in the loop by explaining the proposed changes in everyday language and asking for confirmation before applying them. Tests using the nuPlan autonomous-driving simulator showed that the system could successfully adapt driving behaviour to eight different passenger prompts, producing smoother driving when comfort was prioritised and higher speeds when urgency was expressed.

The approach is significant because it combines the flexibility and natural-language understanding of LLMs with the predictability of traditional autonomous-driving controllers. Rather than relying on an LLM for real-time driving decisions—which can suffer from latency, hallucinations and a lack of performance guarantees—the researchers use GPT-4o-mini only to translate subjective passenger preferences into changes within predefined safe limits. This separation means the underlying motion planner remains responsible for vehicle control, reducing the potential consequences of incorrect LLM outputs. However, researchers note that maintaining safe operating constraints still requires substantial engineering, and the Delft approach does not yet provide mathematically provable safety. Other work is therefore exploring additional verification layers that mathematically check AI-generated driving decisions against traffic rules and predicted road-user behaviour before they are executed.

More information:

https://spectrum.ieee.org/autonomous-vehicles-motion-planner-llm

Flexible Brain Circuits Enable Cognitive Switching

MIT neuroscientists discovered that the brain could reuse the same neural circuits for different cognitive tasks, helping explain how a finite number of neurons supports such a wide range of behaviors. In experiments with mice, researchers examined activity in the prefrontal and parietal cortices while the animals compared auditory tones and remembered information needed to decide. They found that while neurons in the parietal cortex mainly retained sensory information, a group of neurons in the prefrontal cortex could flexibly switch roles, first maintain a sensory memory of a tone and later store the action the animal planned to perform.

The findings support the concept of modular or compositional cognition, in which neural circuits function like reusable building blocks rather than dedicated systems for individual tasks. By applying the same computational mechanisms to different kinds of information (such as sensory inputs or motor plans) the brain can efficiently combine and repurpose neural resources to produce flexible behavior and learn new tasks without constructing entirely new circuits. The researchers plan to test this mechanism further by selectively inhibiting these modules during different stages of tasks to determine their causal role in cognitive flexibility.

More information:

https://news.mit.edu/2026/flexible-brain-circuits-can-switch-between-different-tasks-0817

AI-Designed Viruses Raise Safety Concerns

Scientists at Stanford University have created what are described as the first functional viruses whose genomes were designed by artificial intelligence. Researchers used genome language models called Evo1 and Evo2, trained on genetic information from about 2 million bacteriophages, viruses that infect bacteria, to generate thousands of candidate viral genomes. Nearly 300 were synthesised in the laboratory, producing 16 viable bacteriophages. A mixture of these AI-designed viruses successfully attacked strains of E. coli that had developed resistance to naturally occurring phages, demonstrating potential for developing more adaptable phage therapies against antibiotic-resistant bacterial infections.

The breakthrough also raises significant biosafety and biosecurity concerns because it demonstrates that generative AI can design complete, biologically functional viral genomes. The researchers deliberately excluded genetic data from viruses that infect humans, animals and plants, but experts warn that similar approaches applied to pathogenic organisms could create novel biological threats. Biosecurity specialists therefore argue that governance needs to develop alongside the technology, combining controls on AI-model development and access with responsible research oversight, DNA-synthesis screening, laboratory biosafety and restrictions on potentially dangerous genome production.

More information:

https://www.theguardian.com/science/2026/aug/06/safety-fears-as-scientists-make-first-viruses-designed-by-ai

Teaching Drones to Play Tag with AI

Researchers at Sandia National Laboratories are using reinforcement learning (RL) to teach autonomous drones how to cooperate and make rapid decisions in a strategic version of tag. In the experiment, two evader drones attempt to reach a protected base while two pursuer drones try to intercept them. The evaders learn through simulated rewards and penalties to coordinate their movements, including strategies that can cause the pursuers to interfere or collide with one another. Unlike the RL-trained evaders, the pursuers use a conventional proportional-navigation algorithm. The research demonstrates how multi-agent reinforcement learning can produce flexible cooperative behaviours without engineers having to explicitly program every possible manoeuvre.

The project is overcoming the simulation-to-reality gap: strategies that perform well in simulation can fail on physical drones because of factors such as battery depletion, aerodynamic interactions, communication delays and other unmodelled dynamics. Sandia therefore tests the algorithms on small, inexpensive quadrotors in its CAMINO facility, using high-precision motion capture to evaluate and refine their behaviour before moving toward more costly systems. Beyond the game itself, the work could contribute to autonomous systems capable of pursuit, evasion and coordinated swarm behaviour in rapidly changing environments, including potential applications for protecting critical infrastructure against hostile drones and other national-security scenarios.

More information:

https://www.sandia.gov/labnews/2026/08/13/teaching-drones-to-play-tag/

Humanoid Robots Take on Real-World Firefighting

At the 2^nd World Humanoid Robot Games (WHRG) in Beijing, 23 teams competed in a realistic simulated firefighting challenge aimed at testing whether humanoid robots can progress from impressive demonstrations to practical work in hazardous environments. Held at an actual fire brigade, the competition required robots to complete three tasks within 30 minutes: identify hazardous materials, locate and close open valves, and find and operate a fire extinguisher. The event exposed significant challenges associated with real-world deployment, including rain, changing lighting, object recognition, manipulation accuracy and movement speed. Of the 12 teams competing on Sunday, only three completed the full challenge, highlighting the gap between laboratory performance and reliable operation in unpredictable environments.

Despite these limitations, organisers and participating teams viewed failures as valuable opportunities for improving humanoid robotics. Data from unsuccessful object recognition, manipulation and motion-planning attempts can be used to refine algorithms and train AI models. The competition also demonstrated emerging approaches such as VR-based teleoperation, multi-jointed robotic hands and specialised actuators combining precision with strength. The event reflects China's broader effort to move humanoid robots and embodied AI into practical applications, particularly dangerous tasks such as emergency response. Ultimately, the objective is not necessarily to replace firefighters, but to develop robots capable of working alongside humans and undertaking operations where human exposure would be risky.

More information:

https://www.globaltimes.cn/page/202608/1368322.shtml

AI Assistant for Art Provenance

Researchers are exploring how AI can help solve one of the art world’s most difficult problems: establishing an artwork’s provenance, or documented history of ownership and attribution. The NPR report describes efforts to develop an AI-based provenance assistant capable of searching and connecting information scattered across archives, auction catalogues, museum records, historical documents and other sources. Rather than replacing provenance researchers, the technology is intended to accelerate the labor-intensive process of identifying names, dates, transactions and relationships that may reveal where an artwork has been and who owned it. This could be particularly valuable for works whose histories contain significant gaps or whose ownership changed during periods such as war, displacement or political upheaval.

At the same time, the approach highlights the limitations of using AI for sensitive art-historical research. Provenance evidence can be incomplete, contradictory or ambiguous, meaning that an AI-generated connection cannot automatically be treated as historical fact. Human scholars still need to examine original records, assess the reliability of sources and determine whether proposed connections are credible. The broader promise, therefore, lies in using AI as a research assistant rather than an authority: a tool capable of navigating enormous collections of digitised information and suggesting leads that experts can subsequently verify. Such systems could ultimately make provenance investigations faster and more systematic while helping museums, collectors and researchers address questions of authenticity, ownership and potentially contested cultural property.

More information:

https://www.npr.org/2026/08/06/nx-s1-5922729/ai-art-provenance-assistant