Liquid Neural Networks: Enabling Robots to "Self-Learn" in Dynamic Environments
On July 23, 2026, the School of Computer Science and Engineering at Nanyang Technological University (NTU), Singapore, in collaboration with the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), released the latest research progress on Liquid Neural Networks (LNN). The results demonstrated LNN's adaptive capability in real-time dynamic environments: when robots walk on unknown terrain, they can adjust their gait in real time without pre-training data, consuming only 30% of the energy of traditional deep neural networks while improving decision response speed by over 50%.
LNN differs from traditional deep learning models. Traditional models have fixed parameters after training, while LNN adopts a "liquid" connection mechanism where the connection weights between neurons dynamically adjust as input data changes. This feature makes LNN perform well in scenarios with limited data and dynamic environments. In experiments, the team had robots walk autonomously on different terrains like sand, gravel, and slopes using an LNN controller. The robots completed adaptive adjustments in just 40 seconds, whereas traditional models required over 5,000 training samples to achieve similar performance.
Professor Lin Hongyu, co-leader of the study and director of the NTU AI and Robotics Research Center, said: "Liquid Neural Networks will shift robots from a 'rote memorization' AI model to a 'just-in-time learning' model. This is crucial for applications like industrial automation, home services, and disaster rescue where all scenarios cannot be predicted."
Technical Details: Dynamic Neurons and Weight Plasticity
The core innovation of LNN lies in its neuron structure. Traditional deep networks use fixed weighted connections, while each LNN neuron contains a set of differential equations that govern the temporal dynamics of input signals. This allows the network to process information at different time scales, similar to synaptic plasticity in biological brains. The team used a training algorithm called "Causal Liquidity," enabling the network to quickly adapt to new environments without forgetting previously learned critical knowledge.
Professor Lin noted that LNN has an order of magnitude fewer parameters than comparable CNN or Transformer models. For example, an LNN controlling robot motion requires only about 86 neurons, whereas traditional deep networks need thousands. This significantly reduces computing costs, allowing LNN to run in real time on edge devices without relying on cloud computing.
Currently, the NTU team has developed an LNN-based Robot Operating System (ROS) plugin, open to global developers. The plugin supports mainstream robot platforms such as Fischertechnik and UBTECH.
Industry Applications: Warehousing, Healthcare, and Home Services Lead
The characteristics of LNN give it disruptive potential across multiple industries. In warehouse logistics, Singapore's YCH Group partnered with NTU to deploy LNN-based autonomous handling robots at its Yishun warehouse. These robots adapt to new shelf positions and routes without reprogramming even when warehouse layouts change frequently. According to YCH Group CTO Jeffrey Tan, LNN robots improved picking efficiency by 35% while reducing maintenance costs by 80%.
In healthcare, Singapore General Hospital is testing LNN-controlled exoskeleton robots to assist stroke patients in rehabilitation. The robots dynamically adjust assistance based on real-time EMG signals and gait changes. Hospital representatives said early trials showed patients recovered 20% faster than with traditional programmed exoskeletons.
The home service robot market also sees a new entrant. Startup BotXite plans to launch the first LNN-based household vacuum robot in Q4 2026. The robot can "remember" changes in home environments, such as new furniture or temporarily moved items, and autonomously adjust cleaning strategies based on floor dirt levels.
H2 2026: LNN Commercialization Accelerates
As LNN technology matures, industrial explosion is expected in H2 2026. The Singapore Economic Development Board (EDB) has listed LNN as a key support direction in its National AI Strategy 2.0. Chen Weiming, Director of EDB's Digital Industry Division, said LNN is "truly energy-saving AI" that helps Singapore achieve its 2030 AI computing carbon neutrality goal.
Capital markets have also keenly sensed the trend. In July 2026, Singapore-based VC Jungle Ventures led a $50 million Series B round for LNN chip startup BrainChip. The company's LNN-specific NPU (Neural Processing Unit) chip uses a compute-in-memory architecture, achieving 10 times the energy efficiency of NVIDIA A100 GPU. BrainChip plans to mass-produce the chip by end of 2026 and has signed letters of intent with multiple robot ODM manufacturers.
Meanwhile, the open-source community is also active. The "LiquidNet" open-source project initiated by NTU has garnered over 20,000 stars on GitHub, including pre-trained models, robot control examples, and transfer learning toolkits. Developers can run LNN models on Raspberry Pi or Jetson Nano, greatly lowering the barrier to experimentation.
Expert Views and Risk Warnings
Despite its bright prospects, LNN still faces challenges. MIT CSAIL researcher Susanne S. pointed out that LNN's performance is still inferior to large-scale Transformer models in extremely complex high-dimensional visual tasks. In addition, its lack of theoretical interpretability may hinder application in high-risk fields such as healthcare.
Independent AI analyst and former National University of Singapore researcher Li Ming said: "LNN's advantages in robot control are clear, but it is not a panacea. It is more suitable for scenarios requiring continuous adaptation rather than one-time pattern recognition. For investors, they should focus on companies that closely integrate LNN with specific application scenarios, rather than technology-only startups."
Overall, LNN is transitioning from lab to industrial applications, expected to become a mainstream solution for robot intelligent control in H2 2026. For Singapore and the global AI industry, this is an excellent opportunity to deepen niche tracks and improve ecosystem collaboration.
(Reported by Wang Zhe, Lion City Financial Network)
