Andi Dharmawan

1. The rise of autonomous vehicles has been a pivotal moment in the advancement of transportation, with the focus on providing better safety, efficiency, and integrated mobility solutions (Lee, Par...

An indoor navigation system for an autonomous mobile robot was developed using LiDAR-based perception and multi-sensor fusion. The system combines 2D LiDAR, inertial measurement unit (IMU), and wheel encoder measurements within a simultaneous localization and mapping (SLAM) framework to support real-time localization, while the ROS2 Nav2 stack manages global path planning and local obstacle avoidance through A*-based planning and costmap-driven control. Evaluation in a warehouse-like environment showed that the robot maintained stable localization with low drift and completed autonomous navigation missions with a success rate of 93.33%. During operation, the robot was able to avoid static obstacles consistently and adjust its trajectory in response to simple dynamic obstacles through online replanning. These results indicate that the proposed system is suitable for practical indoor logistics scenarios requiring reliable navigation in structured environments. At the same time, the findings suggest the need for further improvement to handle environments with higher dynamics and denser obstacle configurations.

Although computational resources on robots are often limited, real-time, accurate computation of trigonometric functions is essential in robot manipulators, particularly for forward and inverse kinematics, dynamic analysis, trajectory planning, and motion control. The LUT method requires a large number of LUTs to improve accuracy. The accuracy of the CORDIC method is highly dependent on the number of computational latencies, which affects the computation speed. This paper combines two general approaches for computing trigonometric functions on robot manipulators that improve accuracy without increasing resource utilization and computational latencies. The design uses a 10-bit format (0.125° input resolution and 2-10 output precision) and is implemented in VHDL on a Xilinx Artix-7 XC7A100T-CSG324 FPGA. Compared with a CORDIC-only baseline, the maximum absolute error is reduced from 0.083007812 to 0.009801151 for sine and from 0.079101563 to 0.008901377 for cosine, while MSE drops from 2.4031×10-4 and 2.32974×10-4 to 5.87754×10-6 and 5.87862×10-6, respectively. The hybrid core also reduces slice usage from 81 to 69 and shortens computation time from 35.271 ns to 30.627 ns, making it suitable for resource-constrained real-time robotic control.

Humanoid robots, designed to resemble human structure with a head, arms, and legs, stand out among various robot types due to their ability to interact with human-designed environments. Their kinematic structure, composed of links connected by joints, enables humanoid robots to perform tasks such as walking on uneven terrain, lifting objects, and opening doors. Walking is a critical function for humanoid robots, but it shifts their center of mass (CoM), which can lead to instability and falls. The inverted pendulum model is a widely used approach to humanoid robot walking, particularly on uneven surfaces, enabling CoM projection and balance adjustment. However, disturbances from floor unevenness, mechanical issues, and other factors can still cause the robot to fall. This underscores the need for effective control systems to stabilize movement and maintain balance. The Linear Quadratic Regulator (LQR) is a suitable control approach for humanoid robots, particularly in multiple-input, multiple-output systems. It stabilizes the robot while managing computational load, generating complex humanoid movements through linear approximations. To address uncertainties and adapt control gains, fuzzy control can be integrated, thereby enabling smoother transitions and improved disturbance handling. When walking on uneven terrain, a humanoid robot's body orientation may shift, causing the CoM to exceed tolerance limits and leading to falls. Thus, a control system that adjusts the walking pattern to maintain CoM stability is essential. In conclusion, humanoid robots, with their human-like structure, can perform complex tasks in human environments. Control systems, such as LQR and fuzzy control, are critical for maintaining balance and stability, even under challenging conditions, thereby enabling humanoid robots to keep their trajectory and prevent falls. ABSTRAK: Robot seperti manusia (humanoid), direka menyerupai struktur manusia dengan kepala, lengan, dan kaki, menonjol dalam pelbagai jenis robot kerana keupayaannya berinteraksi dengan persekitaran yang direka oleh manusia. Struktur kinematiknya, terdiri daripada pautan yang dihubungkan oleh sendi, membolehkan robot humanoid melakukan tugas seperti berjalan di permukaan tidak rata, mengangkat objek, dan membuka pintu. Berjalan adalah fungsi penting robot humanoid, tetapi pergerakan mengubah pusat jisim (CoM), menyebabkan ketidakstabilan dan jatuh. Model bandul terbalik adalah pendekatan biasa digunakan bagi robot humanoid berjalan, terutama pada permukaan tidak rata, membolehkan unjuran CoM dan pelarasan keseimbangan. Namun, gangguan ketidaksamaan permukaan lantai, masalah mekanikal, dan faktor lain masih boleh menyebabkan robot jatuh. Ini menunjukkan keperluan pada sistem kawalan berkesan yang menstabilkan pergerakan dan mengekalkan keseimbangan. Regulator Kuadratik Linear (LQR) adalah pendekatan kawalan sesuai untuk robot humanoid, terutama dalam sistem masukan dan keluaran berganda. Ia menstabilkan robot sambil menguruskan beban pengiraan, menghasilkan pergerakan humanoid yang kompleks melalui pendekatan linear. Bagi menangani ketidakpastian dan menyesuai keuntungan kawalan, kawalan kabur boleh diintegrasikan, menjadikan peralihan lebih lancar dan pengendalian gangguan lebih baik. Apabila berjalan di permukaan tidak rata, orientasi badan robot humanoid mungkin berubah, menyebabkan CoM bergerak melebihi had toleransi, yang mengakibatkan jatuh. Oleh itu, sistem kawalan penyesuaian corak berjalan bagi mengekalkan kestabilan CoM adalah penting. Kesimpulannya, robot humanoid, dengan struktur seperti manusia, dapat melakukan tugas kompleks dalam persekitaran manusia. Sistem kawalan seperti LQR dan kawalan kabur adalah penting bagi memastikan keseimbangan dan kestabilan, walaupun dalam keadaan mencabar, membolehkan robot humanoid mengekalkan trajektori mereka dan mencegah jatuh.

This paper presents research and development of an Autonomous Underwater Vehicle (AUV), named ORCA, to perform underwater missions independently without human intervention. ORCA plays a vital role in challenges that test robots' abilities in navigation, exploration, and interaction with the aquatic environment. Using advanced design tools, the AUV is meticulously designed in Inventor and manufactured via CNC machining, laser cutting, and 3D printing. We concentrate on the vehicle's design, manufacturing processes, control systems, PID controllers, and vision systems. Subsequently, the research and development effort is expanded to incorporate critical functionalities, including environmental perception, object detection, deep learning algorithms, and path-planning strategies. The culmination of this research has produced an AUV capable of autonomous underwater navigation, effective obstacle avoidance, efficient object detection, and precise payload manipulation using a gripper mechanism. ORCA has a comprehensive sensor suite comprising a BNO055 IMU, an MS5803-14BA pressure sensor for depth measurement, and a Logitech C525 camera for image processing. The system integration not only enhances the vehicle's operational capabilities but also represents a significant advancement in underwater robotics. ABSTRAK: Kajian ini membentangkan penyelidikan dan pembangunan Kenderaan dalam Air Berautonomi (AUV) yang dikenali sebagai ORCA, dibangunkan bagi tujuan misi bawah air secara autonomi tanpa sebarang campur tangan manusia. ORCA memainkan peranan penting dalam menghadapi cabaran menguji kebolehan robot dalam aspek navigasi, penerokaan, serta interaksi dengan persekitaran akuatik. Dengan memanfaatkan alat reka bentuk yang canggih, AUV ini direka bentuk dengan teliti menggunakan perisian Inventor dan dihasilkan melalui teknik pemesinan CNC, pemotongan laser, dan percetakan 3D. Penekanan utama dalam penyelidikan ini adalah pada reka bentuk kenderaan, proses pembuatan, sistem kawalan, pengawal PID, serta sistem penglihatan. Selain itu, penyelidikan dan pembangunan ini turut diperluaskan bagi menyertakan fungsi-fungsi kritikal seperti persepsi persekitaran, pengesanan objek, algoritma pembelajaran mendalam, dan strategi perancangan laluan. Dapatan kajian ini telah menghasilkan AUV yang mampu melaksanakan navigasi bawah air secara autonomi, menghindari halangan dengan berkesan, mengesan objek dengan cekap, serta mengendali beban dengan tepat menggunakan mekanisme pemegang. ORCA dilengkapi dengan suit pengesan yang komprehensif, termasuk pengesan IMU BNO055, pengesan tekanan MS5803-14BA bagi pengukuran kedalaman, serta kamera Logitech C525 bagi pemprosesan visual. Integrasi sistem ini bukan sahaja meningkatkan keupayaan operasi kenderaan tetapi juga mewakili satu langkah penting dalam perkembangan robotik bawah air.

Achieving precise semantic segmentation is essential for enabling real-time perception in autonomous systems, yet leading approaches typically require substantial annotated data and powerful hardware, restricting their use on devices with limited resources. This work introduces an efficient segmentation framework that integrates pseudo-label refinement, knowledge distillation, and entropy-based confidence filtering to train compact student networks suitable for edge deployment. High-quality pseudo-labels are first produced by a robust teacher network, then further improved using a dense conditional random field to boost spatial consistency. An entropy-based selection mechanism removes unreliable predictions, ensuring that only the most trustworthy labels guide the student model's training. The use of knowledge distillation effectively transfers detailed semantic understanding from the teacher to the student, enhancing accuracy without added computational overhead. Experimental results with multiple EfficientNet backbones reveal that this pipeline improves segmentation accuracy and output clarity, while also supporting real-time or near real-time inference on CPUs with limited processing power. Extensive ablation and qualitative studies further confirm the method's robustness and flexibility for real-world edge applications.

Humanoid robots are designed to replicate human activities, including tasks in hazardous environments. However, maintaining balance during backward walking remains a significant challenge due to center of mass (CoM) shifts beyond the support polygon and limited knee joint motion. This study proposes a control strategy that integrates a linear quadratic regulator (LQR) with optimized walking patterns to enhance dynamic stability. The approach combines LQR-based control with CoM trajectory planning to ensure safe and stable backward walking. The methodology includes inverse kinematics for generating walking patterns and the use of Inertial Measurement Unit (IMU) sensors to estimate the CoM trajectory. LQR parameters were tuned through simulation to improve responsiveness to disturbances. Evaluation metrics focused on CoM deviation, rise time, settling time, and overshoot. Experimental results demonstrate that the proposed LQR system effectively maintains the CoM within 5% of the support polygon boundary. The system achieved rise times under one second and settling times below two seconds, while minimizing pitch and roll overshoots. Compared to proportional control, the proposed method significantly improves stability and reduces the risk of falling. This research advances control strategies for humanoid robots, contributing to improved mobility and operational safety. Moreover, it supports Sustainable Development Goal (SDG) 9 by promoting innovation in intelligent robotic systems that can assist in complex or high-risk environments.

Humanoid robots are designed to replicate human activities, including tasks in hazardous environments. However, maintaining balance during backward walking remains a significant challenge due to center of mass (CoM) shifts beyond the support polygon and limited knee joint motion. This study proposes a control strategy that integrates a linear quadratic regulator (LQR) with optimized walking patterns to enhance dynamic stability. The approach combines LQR-based control with CoM trajectory planning to ensure safe and stable backward walking. The methodology includes inverse kinematics for generating walking patterns and the use of Inertial Measurement Unit (IMU) sensors to estimate the CoM trajectory. LQR parameters were tuned through simulation to improve responsiveness to disturbances. Evaluation metrics focused on CoM deviation, rise time, settling time, and overshoot. Experimental results demonstrate that the proposed LQR system effectively maintains the CoM within 5% of the support polygon boundary. The system achieved rise times under one second and settling times below two seconds, while minimizing pitch and roll overshoots. Compared to proportional control, the proposed method significantly improves stability and reduces the risk of falling. This research advances control strategies for humanoid robots, contributing to improved mobility and operational safety. Moreover, it supports Sustainable Development Goal (SDG) 9 by promoting innovation in intelligent robotic systems that can assist in complex or high-risk environments.

Achieving precise semantic segmentation is essential for enabling real-time perception in autonomous systems, yet leading approaches typically require substantial annotated data and powerful hardware, restricting their use on devices with limited resources. This work introduces an efficient segmentation framework that integrates pseudo-label refinement, knowledge distillation, and entropy-based confidence filtering to train compact student networks suitable for edge deployment. High-quality pseudo-labels are first produced by a robust teacher network, then further improved using a dense conditional random field to boost spatial consistency. An entropy-based selection mechanism removes unreliable predictions, ensuring that only the most trustworthy labels guide the student model's training. The use of knowledge distillation effectively transfers detailed semantic understanding from the teacher to the student, enhancing accuracy without added computational overhead. Experimental results with multiple EfficientNet backbones reveal that this pipeline improves segmentation accuracy and output clarity, while also supporting real-time or near real-time inference on CPUs with limited processing power. Extensive ablation and qualitative studies further confirm the method's robustness and flexibility for real-world edge applications.