Dongsoo Har

Injoon Cho, Praveen Kumar Rajendran, Taeyoung Kim et al.

As the demand for autonomous driving increases, it is paramount to ensure safety. Early accident prediction using deep learning methods for driving safety has recently gained much attention. In this task, early accident prediction and a point prediction of where the drivers should look are determined, with the dashcam video as input. We propose to exploit the double actors and regularized critics (DARC) method, for the first time, on this accident forecasting platform. We derive inspiration from DARC since it is currently a state-of-the-art reinforcement learning (RL) model on continuous action space suitable for accident anticipation. Results show that by utilizing DARC, we can make predictions 5\% earlier on average while improving in multiple metrics of precision compared to existing methods. The results imply that using our RL-based problem formulation could significantly increase the safety of autonomous driving.

Woohyeon Moon, Bumgeun Park, Sarvar Hussain Nengroo et al.

Recently, as the demand for cleaning robots has steadily increased, therefore household electricity consumption is also increasing. To solve this electricity consumption issue, the problem of efficient path planning for cleaning robot has become important and many studies have been conducted. However, most of them are about moving along a simple path segment, not about the whole path to clean all places. As the emerging deep learning technique, reinforcement learning (RL) has been adopted for cleaning robot. However, the models for RL operate only in a specific cleaning environment, not the various cleaning environment. The problem is that the models have to retrain whenever the cleaning environment changes. To solve this problem, the proximal policy optimization (PPO) algorithm is combined with an efficient path planning that operates in various cleaning environments, using transfer learning (TL), detection nearest cleaned tile, reward shaping, and making elite set methods. The proposed method is validated with an ablation study and comparison with conventional methods such as random and zigzag. The experimental results demonstrate that the proposed method achieves improved training performance and increased convergence speed over the original PPO. And it also demonstrates that this proposed method is better performance than conventional methods (random, zigzag).

Quoc-Vinh Lai-Dang, Jihui Lee, Bumgeun Park et al.

Sensor fusion is critical to perception systems for task domains such as autonomous driving and robotics. Recently, the Transformer integrated with CNN has demonstrated high performance in sensor fusion for various perception tasks. In this work, we introduce a method for fusing data from camera and LiDAR. By employing Transformer modules at multiple resolutions, proposed method effectively combines local and global contextual relationships. The performance of the proposed method is validated by extensive experiments with two adversarial benchmarks with lengthy routes and high-density traffics. The proposed method outperforms previous approaches with the most challenging benchmarks, achieving significantly higher driving and infraction scores. Compared with TransFuser, it achieves 8% and 19% improvement in driving scores for the Longest6 and Town05 Long benchmarks, respectively.

Sangkeum Lee, Sarvar Hussain Nengroo, Hojun Jin et al.

With the growth of smart building applications, occupancy information in residential buildings is becoming more and more significant. In the context of the smart buildings' paradigm, this kind of information is required for a wide range of purposes, including enhancing energy efficiency and occupant comfort. In this study, occupancy detection in residential building is implemented using deep learning based on technical information of electric appliances. To this end, a novel approach of occupancy detection for smart residential building system is proposed. The dataset of electric appliances, sensors, light, and HVAC, which is measured by smart metering system and is collected from 50 households, is used for simulations. To classify the occupancy among datasets, the support vector machine and autoencoder algorithm are used. Confusion matrix is utilized for accuracy, precision, recall, and F1 to demonstrate the comparative performance of the proposed method in occupancy detection. The proposed algorithm achieves occupancy detection using technical information of electric appliances by 95.7~98.4%. To validate occupancy detection data, principal component analysis and the t-distributed stochastic neighbor embedding (t-SNE) algorithm are employed. Power consumption with renewable energy system is reduced to 11.1~13.1% in smart buildings by using occupancy detection.

Sangkeum Lee, Sarvar Hussain Nengroo, Hojun Jin et al.

In replacing fossil fuels with renewable energy resources for carbon neutrality, the unbalanced resource production of intermittent wind and photovoltaic (PV) power is a critical issue for peer-to-peer (P2P) power trading. To address this issue, a reinforcement learning (RL) technique is introduced in this paper. For RL, a graph convolutional network (GCN) and a bi-directional long short-term memory (Bi-LSTM) network are jointly applied to P2P power trading between nanogrid clusters, based on cooperative game theory. The flexible and reliable DC nanogrid is suitable for integrating renewable energy for a distribution system. Each local nanogrid cluster takes the position of prosumer, focusing on power production and consumption simultaneously. For the power management of nanogrid cluster, multi-objective optimization is applied to each local nanogrid cluster with the Internet of Things (IoT) technology. Charging/discharging of an electric vehicle (EV) is executed considering the intermittent characteristics of wind and PV power production. RL algorithms, such as GCN- convolutional neural network (CNN) layers for deep Q-learning network (DQN), GCN-LSTM layers for deep recurrent Q-learning network (DRQN), GCN-Bi-LSTM layers for DRQN, and GCN-Bi-LSTM layers for proximal policy optimization (PPO), are used for simulations. Consequently, the cooperative P2P power trading system maximizes the profit by considering the time of use (ToU) tariff-based electricity cost and the system marginal price (SMP), and minimizes the amount of grid power consumption. Power management of nanogrid clusters with P2P power trading is simulated on a distribution test feeder in real time, and the proposed GCN-Bi-LSTM-PPO technique achieving the lowest electricity cost among the RL algorithms used for comparison reduces the electricity cost by 36.7%, averaging over nanogrid clusters.

Sumit Mishra, Medhavi Mishra, Taeyoung Kim et al.

Road infrastructure can affect the occurrence of road accidents. Therefore, identifying roadway features with high accident probability is crucial. Here, we introduce image inpainting that can assist authorities in achieving safe roadway design with minimal intervention in the current roadway structure. Image inpainting is based on inpainting safe roadway elements in a roadway image, replacing accident-prone (AP) features by using a diffusion model. After object-level segmentation, the AP features identified by the properties of accident hotspots are masked by a human operator and safe roadway elements are inpainted. With only an average time of 2 min for image inpainting, the likelihood of an image being classified as an accident hotspot drops by an average of 11.85%. In addition, safe urban spaces can be designed considering human factors of commuters such as gaze saliency. Considering this, we introduce saliency enhancement that suggests chrominance alteration for a safe road view.

Bumgeun Park, Taeyoung Kim, Woohyeon Moon et al.

Training agents via off-policy deep reinforcement learning (RL) requires a large memory, named replay memory, that stores past experiences used for learning. These experiences are sampled, uniformly or non-uniformly, to create the batches used for training. When calculating the loss function, off-policy algorithms assume that all samples are of the same importance. In this paper, we hypothesize that training can be enhanced by assigning different importance for each experience based on their temporal-difference (TD) error directly in the training objective. We propose a novel method that introduces a weighting factor for each experience when calculating the loss function at the learning stage. In addition to improving convergence speed when used with uniform sampling, the method can be combined with prioritization methods for non-uniform sampling. Combining the proposed method with prioritization methods improves sampling efficiency while increasing the performance of TD-based off-policy RL algorithms. The effectiveness of the proposed method is demonstrated by experiments in six environments of the OpenAI Gym suite. The experimental results demonstrate that the proposed method achieves a 33%~76% reduction of convergence speed in three environments and an 11% increase in returns and a 3%~10% increase in success rate for other three environments.

Bumgeun Park, Taeyoung Kim, Quoc-Vinh Lai-Dang et al.

Efficient exploration for an agent is challenging in reinforcement learning (RL). In this paper, a novel actor-critic framework namely virtual action actor-critic (VAAC), is proposed to address the challenge of efficient exploration in RL. This work is inspired by humans' ability to imagine the potential outcomes of their actions without actually taking them. In order to emulate this ability, VAAC introduces a new actor called virtual actor (VA), alongside the conventional actor-critic framework. Unlike the conventional actor, the VA takes the virtual action to anticipate the next state without interacting with the environment. With the virtual policy following a Gaussian distribution, the VA is trained to maximize the anticipated novelty of the subsequent state resulting from a virtual action. If any next state resulting from available actions does not exhibit high anticipated novelty, training the VA leads to an increase in the virtual policy entropy. Hence, high virtual policy entropy represents that there is no room for exploration. The proposed VAAC aims to maximize a modified Q function, which combines cumulative rewards and the negative sum of virtual policy entropy. Experimental results show that the VAAC improves the exploration performance compared to existing algorithms.

Woohyeon Moon, Taeyoung Kim, Bumgeun Park et al.

Transformer is a state-of-the-art model in the field of natural language processing (NLP). Current NLP models primarily increase the number of transformers to improve processing performance. However, this technique requires a lot of training resources such as computing capacity. In this paper, a novel structure of Transformer is proposed. It is featured by full layer normalization, weighted residual connection, positional encoding exploiting reinforcement learning, and zero masked self-attention. The proposed Transformer model, which is called Enhanced Transformer, is validated by the bilingual evaluation understudy (BLEU) score obtained with the Multi30k translation dataset. As a result, the Enhanced Transformer achieves 202.96% higher BLEU score as compared to the original transformer with the translation dataset.

Bumgeun Park, Jihui Lee, Taeyoung Kim et al.

This paper presents a technique to train a robot to perform kick-motion in AI soccer by using reinforcement learning (RL). In RL, an agent interacts with an environment and learns to choose an action in a state at each step. When training RL algorithms, a problem called the curse of dimensionality (COD) can occur if the dimension of the state is high and the number of training data is low. The COD often causes degraded performance of RL models. In the situation of the robot kicking the ball, as the ball approaches the robot, the robot chooses the action based on the information obtained from the soccer field. In order not to suffer COD, the training data, which are experiences in the case of RL, should be collected evenly from all areas of the soccer field over (theoretically infinite) time. In this paper, we attempt to use the relative coordinate system (RCS) as the state for training kick-motion of robot agent, instead of using the absolute coordinate system (ACS). Using the RCS eliminates the necessity for the agent to know all the (state) information of entire soccer field and reduces the dimension of the state that the agent needs to know to perform kick-motion, and consequently alleviates COD. The training based on the RCS is performed with the widely used Deep Q-network (DQN) and tested in the AI Soccer environment implemented with Webots simulation software.

Praveen Kumar Rajendran, Quoc-Vinh Lai-Dang, Luiz Felipe Vecchietti et al.

Identifying the camera pose for a given image is a challenging problem with applications in robotics, autonomous vehicles, and augmented/virtual reality. Lately, learning-based methods have shown to be effective for absolute camera pose estimation. However, these methods are not accurate when generalizing to different domains. In this paper, a domain adaptive training framework for absolute pose regression is introduced. In the proposed framework, the scene image is augmented for different domains by using generative methods to train parallel branches using Barlow Twins objective. The parallel branches leverage a lightweight CNN-based absolute pose regressor architecture. Further, the efficacy of incorporating spatial and channel-wise attention in the regression head for rotation prediction is investigated. Our method is evaluated with two datasets, Cambridge landmarks and 7Scenes. The results demonstrate that, even with using roughly 24 times fewer FLOPs, 12 times fewer activations, and 5 times fewer parameters than MS-Transformer, our approach outperforms all the CNN-based architectures and achieves performance comparable to transformer-based architectures. Our method ranks 2nd and 4th with the Cambridge Landmarks and 7Scenes datasets, respectively. In addition, for augmented domains not encountered during training, our approach significantly outperforms the MS-transformer. Furthermore, it is shown that our domain adaptive framework achieves better performance than the single branch model trained with the identical CNN backbone with all instances of the unseen distribution.

Taeyoung Kim, Dongsoo Har

In multi-goal reinforcement learning with a sparse binary reward, training agents is particularly challenging, due to a lack of successful experiences. To solve this problem, hindsight experience replay (HER) generates successful experiences even from unsuccessful ones. However, generating successful experiences from uniformly sampled ones is not an efficient process. In this paper, the impact of exploiting the property of achieved goals in generating successful experiences is investigated and a novel cluster-based sampling strategy is proposed. The proposed sampling strategy groups episodes with different achieved goals by using a cluster model and samples experiences in the manner of HER to create the training batch. The proposed method is validated by experiments with three robotic control tasks of the OpenAI Gym. The results of experiments demonstrate that the proposed method is substantially sample efficient and achieves better performance than baseline approaches.

Andres Saurez, Neha Sengar, Dongsoo Har

Circuit discovery and activation steering in transformers have developed as separate research threads, yet both operate on the same representational space. Are they two views of the same underlying structure? We show they follow a single geometric principle: answer tokens, processed in isolation, encode the directions that would produce them. This Circuit Fingerprint hypothesis enables circuit discovery without gradients or causal intervention -- recovering comparable structure to gradient-based methods through geometric alignment alone. We validate this on standard benchmarks (IOI, SVA, MCQA) across four model families, achieving circuit discovery performance comparable to gradient-based methods. The same directions that identify circuit components also enable controlled steering -- achieving 69.8\% emotion classification accuracy versus 53.1\% for instruction prompting while preserving factual accuracy. Beyond method development, this read-write duality reveals that transformer circuits are fundamentally geometric structures: interpretability and controllability are two facets of the same object.

Yousung Lee, Dongsoo Har

Dynamic head pruning in Vision Transformers (ViTs) improves efficiency by removing redundant attention heads, but existing pruning policies are often difficult to interpret and control. In this work, we propose a novel framework by integrating Sparse Autoencoders (SAEs) with dynamic pruning, leveraging their ability to disentangle dense embeddings into interpretable and controllable sparse latents. Specifically, we train an SAE on the final-layer residual embedding of the ViT and amplify the sparse latents with different strategies to alter pruning decisions. Among them, per-class steering reveals compact, class-specific head subsets that preserve accuracy. For example, bowl improves accuracy (76% to 82%) while reducing head usage (0.72 to 0.33) via heads h2 and h5. These results show that sparse latent features enable class-specific control of dynamic pruning, effectively bridging pruning efficiency and mechanistic interpretability in ViTs.

Andres Saurez, Yousung Lee, Dongsoo Har

Linear probes and sparse autoencoders consistently recover meaningful structure from transformer representations -- yet why should such simple methods succeed in deep, nonlinear systems? We show this is not merely an empirical regularity but a consequence of architectural necessity: transformers communicate information through linear interfaces (attention OV circuits, unembedding matrices), and any semantic feature decoded through such an interface must occupy a context-invariant linear subspace. We formalize this as the \emph{Invariant Subspace Necessity} theorem and derive the \emph{Self-Reference Property}: tokens directly provide the geometric direction for their associated features, enabling zero-shot identification of semantic structure without labeled data or learned probes. Empirical validation in eight classification tasks and four model families confirms the alignment between class tokens and semantically related instances. Our framework provides \textbf{a principled architectural explanation} for why linear interpretability methods work, unifying linear probes and sparse autoencoders.

Neha Sengar, Indra Kumari, Jihui Lee et al.

Driver distraction is a principal cause of traffic accidents. In a study conducted by the National Highway Traffic Safety Administration, engaging in activities such as interacting with in-car menus, consuming food or beverages, or engaging in telephonic conversations while operating a vehicle can be significant sources of driver distraction. From this viewpoint, this paper introduces a novel method for detection of driver distraction using multi-view driver action images. The proposed method is a vision transformer-based framework with pose estimation and action inference, namely PoseViNet. The motivation for adding posture information is to enable the transformer to focus more on key features. As a result, the framework is more adept at identifying critical actions. The proposed framework is compared with various state-of-the-art models using SFD3 dataset representing 10 behaviors of drivers. It is found from the comparison that the PoseViNet outperforms these models. The proposed framework is also evaluated with the SynDD1 dataset representing 16 behaviors of driver. As a result, the PoseViNet achieves 97.55% validation accuracy and 90.92% testing accuracy with the challenging dataset.

Seungah Son, Andrez Saurez, Dongsoo Har

While pre-trained language models excel at semantic understanding, they often struggle to capture nuanced affective information critical for affective recognition tasks. To address these limitations, we propose a novel framework for enhancing emotion-aware embeddings in transformer-based models. Our approach introduces a continuous valence-arousal labeling system to guide contrastive learning, which captures subtle and multi-dimensional emotional nuances more effectively. Furthermore, we employ a dynamic token perturbation mechanism, using gradient-based saliency to focus on sentiment-relevant tokens, improving model sensitivity to emotional cues. The experimental results demonstrate that the proposed framework outperforms existing methods, achieving up to 15.5% improvement in the emotion classification benchmark, highlighting the importance of employing continuous labels. This improvement demonstrates that the proposed framework is effective in affective representation learning and enables precise and contextually relevant emotional understanding.

Inkyu Kim, Sangkeum Lee, Haechan Jeong et al.

Satellite communication systems (SCSs) used for tactical purposes require robust security and anti-jamming capabilities, making frequency hopping (FH) a powerful option. However, the current FH systems face challenges due to significant interference from other devices and the considerable path loss inherent in satellite communication. This misalignment leads to inefficient synchronization, crucial for maintaining reliable communication. Traditional methods, such as those employing long short-term memory (LSTM) networks, have made improvements, but they still struggle in dynamic conditions of satellite environments. This paper presents a novel method for synchronizing FH signals in tactical SCSs by combining serial search and reinforcement learning to achieve coarse and fine acquisition, respectively. The mathematical analysis and simulation results demonstrate that the proposed method reduces the average number of hops required for synchronization by 58.17% and mean squared error (MSE) of the uplink hop timing estimation by 76.95%, as compared to the conventional serial search method. Comparing with the early late gate synchronization method based on serial search and use of LSTM network, the average number of hops for synchronization is reduced by 12.24% and the MSE by 18.5%.

Praveen Kumar Rajendran, Sumit Mishra, Luiz Felipe Vecchietti et al.

Relative camera pose estimation, i.e. estimating the translation and rotation vectors using a pair of images taken in different locations, is an important part of systems in augmented reality and robotics. In this paper, we present an end-to-end relative camera pose estimation network using a siamese architecture that is independent of camera parameters. The network is trained using the Cambridge Landmarks data with four individual scene datasets and a dataset combining the four scenes. To improve generalization, we propose a novel two-stage training that alleviates the need of a hyperparameter to balance the translation and rotation loss scale. The proposed method is compared with one-stage training CNN-based methods such as RPNet and RCPNet and demonstrate that the proposed model improves translation vector estimation by 16.11%, 28.88%, and 52.27% on the Kings College, Old Hospital, and St Marys Church scenes, respectively. For proving texture invariance, we investigate the generalization of the proposed method augmenting the datasets to different scene styles, as ablation studies, using generative adversarial networks. Also, we present a qualitative assessment of epipolar lines of our network predictions and ground truth poses.

Sumit Mishra, Praveen Kumar Rajendran, Luiz Felipe Vecchietti et al.

In urban cities, visual information on and along roadways is likely to distract drivers and lead to missing traffic signs and other accident-prone (AP) features. To avoid accidents due to missing these visual cues, this paper proposes a visual notification of AP-features to drivers based on real-time images obtained via dashcam. For this purpose, Google Street View images around accident hotspots (areas of dense accident occurrence) identified by a real-accident dataset are used to train a novel attention module to classify a given urban scene into an accident hotspot or a non-hotspot (area of sparse accident occurrence). The proposed module leverages channel, point, and spatial-wise attention learning on top of different CNN backbones. This leads to better classification results and more certain AP-features with better contextual knowledge when compared with CNN backbones alone. Our proposed module achieves up to 92% classification accuracy. The capability of detecting AP-features by the proposed model were analyzed by a comparative study of three different class activation map (CAM) methods, which were used to inspect specific AP-features causing the classification decision. Outputs of CAM methods were processed by an image processing pipeline to extract only the AP-features that are explainable to drivers and notified using a visual notification system. Range of experiments was performed to prove the efficacy and AP-features of the system. Ablation of the AP-features taking 9.61%, on average, of the total area in each image increased the chance of a given area to be classified as a non-hotspot by up to 21.8%.

Sumit Mishra, Praveen Kumar Rajendran, Dongsoo Har

In urban cities, with increasing acceptability of shared spaces used by pedestrians and personal mobility devices (PMDs), there is need for pragmatic socially ac-ceptable path planning and navigation management policies. Hence, we propose a socially acceptable global route planner and assess the legibility of the resulting global route. Our approach proposed for choosing global route avoids streets penetrating shared spaces and main routes with high probability of dense usage. The experimental study shows that socially acceptable routes can be effectively found with an average of 10 % increment of route length with optimal hyperpa-rameters. This helps PMDs to reach the goal while taking a socially acceptable and safe route with minimal interaction of different PMDs and pedestrians. When PMDs interact with pedestrians and other types of PMDs in shared spaces, mi-cro-mobility simulations are of prime usage for acceptable and safe navigation policy. Social force models being state of the art for pedestrian simulation are cal-ibrated for capturing random movements of pedestrian behavior. Social force model with calibration can imitate the required behavior of PMDs in a pedestrian mix navigation scheme. Based on calibrated models, simulations on shared space links and gate structures are performed to assist policies related to deciding wait-ing and stopping time. Also, based on simulated PMDs interaction with pedestri-ans, location data with finer resolution can be obtained if the resolution of GPS sensor is 0.2 m or less. This will help in formalizing better modelling and hence better micro-mobility policies.

Sangkeum Lee, Hojun Jin, Sarvar Hussain Nengroo et al.

Anomaly detection is concerned with a wide range of applications such as fault detection, system monitoring, and event detection. Identifying anomalies from metering data obtained from smart metering system is a critical task to enhance reliability, stability, and efficiency of the power system. This paper presents an anomaly detection process to find outliers observed in the smart metering system. In the proposed approach, bi-directional long short-term memory (BiLSTM) based autoencoder is used and finds the anomalous data point. It calculates the reconstruction error through autoencoder with the non-anomalous data, and the outliers to be classified as anomalies are separated from the non-anomalous data by predefined threshold. Anomaly detection method based on the BiLSTM autoencoder is tested with the metering data corresponding to 4 types of energy sources electricity/water/heating/hot water collected from 985 households.

Inhwan Kim, Sarvar Hussain Nengroo, Dongsoo Har

This paper presents a technique for navigation of mobile robot with Deep Q-Network (DQN) combined with Gated Recurrent Unit (GRU). The DQN integrated with the GRU allows action skipping for improved navigation performance. This technique aims at efficient navigation of mobile robot such as autonomous parking robot. Framework for reinforcement learning can be applied to the DQN combined with the GRU in a real environment, which can be modeled by the Partially Observable Markov Decision Process (POMDP). By allowing action skipping, the ability of the DQN combined with the GRU in learning key-action can be improved. The proposed algorithm is applied to explore the feasibility of solution in real environment by the ROS-Gazebo simulator, and the simulation results show that the proposed algorithm achieves improved performance in navigation and collision avoidance as compared to the results obtained by DQN alone and DQN combined with GRU without allowing action skipping.

Taeyoung Kim, Luiz Felipe Vecchietti, Kyujin Choi et al.

In multi-agent reinforcement learning, the cooperative learning behavior of agents is very important. In the field of heterogeneous multi-agent reinforcement learning, cooperative behavior among different types of agents in a group is pursued. Learning a joint-action set during centralized training is an attractive way to obtain such cooperative behavior, however, this method brings limited learning performance with heterogeneous agents. To improve the learning performance of heterogeneous agents during centralized training, two-stage heterogeneous centralized training which allows the training of multiple roles of heterogeneous agents is proposed. During training, two training processes are conducted in a series. One of the two stages is to attempt training each agent according to its role, aiming at the maximization of individual role rewards. The other is for training the agents as a whole to make them learn cooperative behaviors while attempting to maximize shared collective rewards, e.g., team rewards. Because these two training processes are conducted in a series in every timestep, agents can learn how to maximize role rewards and team rewards simultaneously. The proposed method is applied to 5 versus 5 AI robot soccer for validation. Simulation results show that the proposed method can train the robots of the robot soccer team effectively, achieving higher role rewards and higher team rewards as compared to other approaches that can be used to solve problems of training cooperative multi-agent.