International Journal of Advanced AI Applications

EchoKG: A Dynamic user Preference Knowledge Graph In-vehicle Dialogue System Based on Ebbinghaus Forgetting Curve

2026-01-21T08:19:14+00:00

With the increasing integration of large language models (LLMs) into intelligent vehicle cockpits, achieving efficient, accurate, and personalized interactions with long-term memory capabilities has become a key challenge. Existing vector retrieval methods suffer from context inflation issues, while static knowledge graphs struggle to capture the time-varying nature of user preferences. This paper proposes the EchoKG framework, which for the first time mathematically models the Ebbinghaus forgetting curve as a dynamic weight mechanism for knowledge graph nodes, enabling the natural decay and reinforcement of user preferences. By introducing memory strength S and last access time, EchoKG dynamically manages the lifecycle of memories. Experimental results on the fully open-source dataset EchoCar-Public demonstrate that compared to MemoryBank, static knowledge graphs, and GPT-4o Memory, EchoKG reduces the average context length by 32%, increases the F1 score for intent recognition by 5.1%, and improves the personalized consistency score by 0.68 points, while maintaining a response latency within 800ms.

Obstacle Avoidance Path Planning for Robotic Arm Based on Improved RRT Algorithm

2026-01-11T02:37:17+00:00

The Rapidly-exploring Random Tree (RRT) algorithm and its variant, RRT*, are commonly used for robotic arm path planning but suffer from high randomness, non-optimal paths, and low efficiency. To address these issues, this paper proposes an improved RRT* algorithm that incorporates a goal-biased sampling strategy and cubic B-spline curve fitting. The method defines and dynamically restricts the search area during tree expansion to improve planning efficiency and goal orientation. Subsequently, cubic B-spline fitting is applied to smooth the path and reduce redundant nodes. Simulation experiments conducted in Python demonstrate that compared to traditional RRT and RRT* algorithms, the proposed approach generates shorter paths with fewer nodes and higher planning success rates, validating its effectiveness for robotic arm obstacle avoidance path planning.

Design of an Adaptive Stair-Climbing Robot Based on Heterogeneous Dual-Core Intelligent Control Technology

2025-12-24T15:38:34+00:00

With the deepening trend of societal aging, the demand for mobile robots in scenarios such as elderly assistance, disability aid, logistics, and rescue is growing. Navigating stairs in complex, unstructured environments has become a key challenge in robotics. Traditional wheeled, tracked, or legged robots suffer from weak adaptability, insufficient stability, or high cost. This paper designs an adaptive stair-climbing robot utilizing a heterogeneous dual-core control architecture built with an STM32H743 microcontroller and a Raspberry Pi 4B. It integrates multiple sensors including an RGB-D camera, an Inertial Measurement Unit (IMU), and encoders. The Raspberry Pi 4B serves as the upper-layer intelligent decision-making core, performing planning and decision-making through fuzzy logic and Model Predictive Control (MPC). The STM32H743 acts as the lower-layer real-time control core, achieving precise execution via PID control. The robot can adapt to stairs with slopes of 30°–45° and step heights of 150–200 mm made of different materials, maintaining a stability margin of no less than 20 mm during climbing. Compared to traditional tracked robots, the stability margin is improved by over 35%. The robot demonstrates good stability and robustness in various stair environments, providing an innovative technical approach for mobile robots in complex terrains.

Research on Bio-inspired Self-balancing Control Based on LIF Network

2026-01-19T03:52:37+00:00

Human balance is a skill gradually established through a sensory-action-feedback loop, relying on repetitive training, trial-and-error mechanisms, and the dynamic plasticity of synaptic connections. In this process, sensory signals are continuously transmitted to the central nervous system, where stable motor paths are formed through learning, enabling action reuse without complex calculations. Inspired by this mechanism, this paper proposes a balance learning method based on brain-like spiking neural networks and dopamine-modulated synaptic plasticity for self-learning control of the classic inverted pendulum system. The method connects the one-hot encoded sensory neuron group with motor neurons and utilizes a reward-driven synaptic weight update mechanism to gradually master the stable control of the inverted pendulum without the need for prior models or training data. Unlike traditional control algorithms such as PID or LQR, this approach features biological realism, strong adaptability, and self-organizing behavior, providing a new perspective on bio-inspired learning strategies for artificial intelligence in continuous control tasks.

Research on Fine-grained Detection Method of Honey Pot Contracts Based on LSTM and Fuzzing

2025-12-28T03:48:46+00:00

Honey pot contract operation code sequences exhibit strong concealment, significantly increasing detection complexity. To address this, this study proposes a fine-grained detection method based on LSTM and Fuzzing. By analyzing frequency differences across operation codes in different honey pot contract types, we calculate their occurrence rates and assign high initial weights to high-frequency operation codes. The weight mechanism is then integrated into the LSTM model to calculate operational code contribution levels and importance scores, enabling extraction of high-scoring critical operation codes. The research employs Fuzzing fuzz testing technology to generate initial test case sets and defines their deconstruction methods. Using case identifiers and functional codes, we validate interaction logic vulnerabilities in honey pot contracts through mutation factor probability matrices. By constructing source code graph structures using critical operation codes and interaction logic vulnerabilities, we update and aggregate vector nodes with global accumulation pooling functions to generate graph-level vectors. These graph-level vectors are then fed into graph attention networks, with cross-entropy loss functions jointly determining honey pot contract types. Test results demonstrate that the proposed method achieves sub-3 false positives for six honey pot contract types, demonstrating high precision in fine-grained detection.