Learning-based Motion Research Articles

Optical Flow with Learning Feature for Deformable Medical Image Registration

Deformable medical image registration plays a vital role in medical image applications, such as placing different temporal images at the same time point or different modality images into the same coordinate system. Various strategies have been developed to satisfy the increasing needs of deformable medical image registration. One popular registration method is estimating the displacement field by computing the optical flow between two images. The motion field (flow field) is computed based on either gray-value or handcrafted descriptors such as the scale-invariant feature transform (SIFT). These methods assume that illumination is constant between images. However, medical images may not always satisfy this assumption. In this study, we propose a metric learning-based motion estimation method called Siamese Flow for deformable medical image registration. We train metric learners using a Siamese network, which produces an image patch descriptor that guarantees a smaller feature distance in two similar anatomical structures and a larger feature distance in two dissimilar anatomical structures. In the proposed registration framework, the flow field is computed based on such features and is close to the real deformation field due to the excellent feature representation ability of the Siamese network. Experimental results demonstrate that the proposed method outperforms the Demons, SIFT Flow, Elastix, and VoxelMorph networks regarding registration accuracy and robustness, particularly with large deformations.

Imitation Learning for Variable Speed Contact Motion for Operation up to Control Bandwidth

Thegeneration of robot motions in the real world is difficult by using conventional controllers alone and requires highly intelligent processing. In this regard, learning-based motion generations are currently being investigated. However, the main issue has been improvements of the adaptability to spatially varying environments, but a variation of the operating speed has not been investigated in detail. In contact-rich tasks, it is especially important to be able to adjust the operating speed because a nonlinear relationship occurs between the operating speed and force (e.g., inertial and frictional forces), and it affects the results of the tasks. Therefore, in this study, we propose a method for generating variable operating speeds while adapting to spatial perturbations in the environment. The proposed method can be adapted to nonlinearities by utilizing a small amount of motion data. We experimentally evaluated the proposed method by erasing a line using an eraser fixed to the tip of the robot as an example of a contact-rich task. Furthermore, the proposed method enables a robot to perform a task faster than a human operator and is capable of operating close to the control bandwidth.

ME‐MADDPG: An efficient learning‐based motion planning method for multiple agents in complex environments

Developing efficient motion policies for multiagents is a challenge in a decentralized dynamic situation, where each agent plans its own paths without knowing the policies of the other agents involved. This paper presents an efficient learning-based motion planning method for multiagent systems. It adopts the framework of multiagent deep deterministic policy gradient (MADDPG) to directly map partially observed information to motion commands for multiple agents. To improve the efficiency of MADDPG in sample utilization, so as to train more brilliant agents that can adapt to more complex environments, a strategy named mixed experience (ME) is introduced to MADDPG, and this has led to our proposed ME-MADDPG algorithm. The novel ME strategy can be embodied into three specific mechanisms: (1) an artificial potential field-based sample generator to produce high-quality samples in the early training stage; (2) a dynamic mixed sampling strategy to mix the training data from different sources with a variable proportion; (3) a delayed learning skill to stabilize the training of the multiple agents. A series of experiments have been conducted to verify the performance of the proposed ME-MADDPG algorithm, and it has been demonstrated that, compared with MADDPG, the proposed algorithm can significantly improve the convergence speed and convergence effect in the training process, and it has also shown better efficiency and better adaptability in complex dynamic environments while it is used for multiagent motion planning applications.

Learning-Based Motion Planning of a 14-DOF Biped Robot on 3D Uneven Terrain Containing a Ditch

In this paper, a new method is proposed that integrates the 3D terrain information, ditch geometry, and biped dynamics for motion planning of a 14-DOF biped robot on 3D terrain containing a 3D ditch. The path planning is modeled as a wavefront propagation in Non-uniform medium represented by the Eikonal equation. A speed function expresses the inhomogeneity of uneven terrain. The Eikonal equation is solved by the Fast marching method (FMM) to obtain the global path. Ten Footstep variables (FSVs) characterize one step of walk on a 3D uneven terrain surface. The hip and foot trajectory parameters (HFTPs) are used to construct cubic spline-based hip and foot trajectories of the biped robot. The optimal value of HFTPs is obtained by a Genetic algorithm by minimization of energy and satisfying the constraint of dynamic balance. A walk-dataset is created that contains the optimal trajectories for different FSVs. The generated walk-dataset was used to train the biped to walk on rough terrain using a feed-forward Neural network for making a real-time estimate of optimal HFTPs. Simulation results validate the effectiveness of the proposed method of ditch crossing on different uneven terrains.

Automatic Inter-Frame Patient Motion Correction for Dynamic Cardiac PET Using Deep Learning

Patient motion during dynamic PET imaging can induce errors in myocardial blood flow (MBF) estimation. Motion correction for dynamic cardiac PET is challenging because the rapid tracer kinetics of 82Rb leads to substantial tracer distribution change across different dynamic frames over time, which can cause difficulties for image registration-based motion correction, particularly for early dynamic frames. In this paper, we developed an automatic deep learning-based motion correction (DeepMC) method for dynamic cardiac PET. In this study we focused on the detection and correction of inter-frame rigid translational motion caused by voluntary body movement and pattern change of respiratory motion. A bidirectional-3D LSTM network was developed to fully utilize both local and nonlocal temporal information in the 4D dynamic image data for motion detection. The network was trained and evaluated over motion-free patient scans with simulated motion so that the motion ground-truths are available, where one million samples based on 65 patient scans were used in training, and 600 samples based on 20 patient scans were used in evaluation. The proposed method was also evaluated using additional 10 patient datasets with real motion. We demonstrated that the proposed DeepMC obtained superior performance compared to conventional registration-based methods and other convolutional neural networks (CNN), in terms of motion estimation and MBF quantification accuracy. Once trained, DeepMC is much faster than the registration-based methods and can be easily integrated into the clinical workflow. In the future work, additional investigation is needed to evaluate this approach in a clinical context with realistic patient motion.

Self-supervised learning-based diffeomorphic non-rigid motion estimation for fast motion-compensated coronary MR angiography

PurposeTo accelerate non-rigid motion corrected coronary MR angiography (CMRA) reconstruction by developing a deep learning based non-rigid motion estimation network and combining this with an efficient implementation of the undersampled motion corrected reconstruction. MethodsUndersampled and respiratory motion corrected CMRA with overall short scans of 5 to 10 min have been recently proposed. However, image reconstruction with this approach remains lengthy, since it relies on several non-rigid image registrations to estimate the respiratory motion and on a subsequent iterative optimization to correct for motion during the undersampled reconstruction. Here we introduce a self-supervised diffeomorphic non-rigid respiratory motion estimation network, DiRespME-net, to speed up respiratory motion estimation. We couple this with an efficient GPU-based implementation of the subsequent motion-corrected iterative reconstruction. DiRespME-net is based on a U-Net architecture, and is trained in a self-supervised fashion, with a loss enforcing image similarity and spatial smoothness of the motion fields. Motion predicted by DiRespME-net was used for GPU-based motion-corrected CMRA in 12 test subjects and final images were compared to those produced by state-of-the-art reconstruction. Vessel sharpness and visible length of the right coronary artery (RCA) and the left anterior descending (LAD) coronary artery were used as metrics of image quality for comparison. ResultsNo statistically significant difference in image quality was found between images reconstructed with the proposed approach (MC:DiRespME-net) and a motion-corrected reconstruction using cubic B-splines (MC:Nifty-reg). Visible vessel length was not significantly different between methods (RCA: MC:Nifty-reg 5.7 ± 1.7 cm vs MC:DiRespME-net 5.8 ± 1.7 cm, P = 0.32; LAD: MC:Nifty-reg 7.0 ± 2.6 cm vs MC:DiRespME-net 6.9 ± 2.7 cm, P = 0.81). Similarly, no statistically significant difference between methods was observed in terms of vessel sharpness (RCA: MC:Nifty-reg 60.3 ± 7.2% vs MC:DiRespME-net 61.0 ± 6.8%, P = 0.19; LAD: MC:Nifty-reg 57.4 ± 7.9% vs MC:DiRespME-net 58.1 ± 7.5%, P = 0.27). The proposed approach achieved a 50-fold reduction in computation time, resulting in a total reconstruction time of approximately 20 s. ConclusionsThe proposed self-supervised learning-based motion corrected reconstruction enables fast motion-corrected CMRA image reconstruction, holding promise for integration in clinical routine.

Learning-based Robust Motion Planning With Guaranteed Stability: A Contraction Theory Approach

This letter presents Learning-based Autonomous Guidance with RObustness and Stability guarantees (LAG-ROS), which provides machine learning-based nonlinear motion planners with formal robustness and stability guarantees, by designing a differential Lyapunov function using contraction theory. LAG-ROS utilizes a neural network to model a robust tracking controller independently of a target trajectory, for which we show that the Euclidean distance between the target and controlled trajectories is exponentially bounded linearly in the learning error, even under the existence of bounded external disturbances. We also present a convex optimization approach that minimizes the steady-state bound of the tracking error to construct the robust control law for neural network training. In numerical simulations, it is demonstrated that the proposed method indeed possesses superior properties of robustness and nonlinear stability resulting from contraction theory, whilst retaining the computational efficiency of existing learning-based motion planners.

Coronary Magnetic Resonance Angiography in Chronic Coronary Syndromes.

Cardiovascular disease is the leading cause of mortality worldwide, with atherosclerotic coronary artery disease (CAD) accounting for the majority of cases. X-ray coronary angiography and computed tomography coronary angiography (CCTA) are the imaging modalities of choice for the assessment of CAD. However, the use of ionising radiation and iodinated contrast agents remain drawbacks. There is therefore a clinical need for an alternative modality for the early identification and longitudinal monitoring of CAD without these associated drawbacks. Coronary magnetic resonance angiography (CMRA) could be a potential alternative for the detection and monitoring of coronary arterial stenosis, without exposing patients to ionising radiation or iodinated contrast agents. Further advantages include its versatility, excellent soft tissue characterisation and suitability for repeat imaging. Despite the early promise of CMRA, widespread clinical utilisation remains limited due to long and unpredictable scan times, onerous scan planning, lower spatial resolution, as well as motion related image quality degradation. The past decade has brought about a resurgence in CMRA technology, with significant leaps in image acceleration, respiratory and cardiac motion estimation and advanced motion corrected or motion-resolved image reconstruction. With the advent of artificial intelligence, great advances are also seen in deep learning-based motion estimation, undersampled and super-resolution reconstruction promising further improvements of CMRA. This has enabled high spatial resolution (1 mm isotropic), 3D whole heart CMRA in a clinically feasible and reliable acquisition time of under 10 min. Furthermore, latest super-resolution image reconstruction approaches which are currently under evaluation promise acquisitions as short as 1 min. In this review, we will explore the recent technological advances that are designed to bring CMRA closer to clinical reality.

Myocardial Function Imaging in Echocardiography Using Deep Learning.

Deformation imaging in echocardiography has been shown to have better diagnostic and prognostic value than conventional anatomical measures such as ejection fraction. However, despite clinical availability and demonstrated efficacy, everyday clinical use remains limited at many hospitals. The reasons are complex, but practical robustness has been questioned, and a large inter-vendor variability has been demonstrated. In this work, we propose a novel deep learning based framework for motion estimation in echocardiography, and use this to fully automate myocardial function imaging. A motion estimator was developed based on a PWC-Net architecture, which achieved an average end point error of (0.06±0.04) mm per frame using simulated data from an open access database, on par or better compared to previously reported state of the art. We further demonstrate unique adaptability to image artifacts such as signal dropouts, made possible using trained models that incorporate relevant image augmentations. Further, a fully automatic pipeline consisting of cardiac view classification, event detection, myocardial segmentation and motion estimation was developed and used to estimate left ventricular longitudinal strain in vivo. The method showed promise by achieving a mean deviation of (-0.7±1.6)% compared to a semi-automatic commercial solution for N=30 patients with relevant disease, within the expected limits of agreement. We thus believe that learning-based motion estimation can facilitate extended use of strain imaging in clinical practice.

REEEC-AGENT: human driver cognition and emotions-inspired rear-end collision avoidance method for autonomous vehicles

Rear-end collision detection and avoidance is one of the most crucial driving tasks of self-driving vehicles. Mathematical models and fuzzy logic-based methods have recently been proposed to improve the effectiveness of the rear-end collision detection and avoidance systems in autonomous vehicles (AVs). However, these methodologies do not tackle real-time object detection and response problems in dense/dynamic road traffic conditions due to their complex computation and decision-making structures. In our previous work, we presented an affective computing-inspired Enhanced Emotion Enabled Cognitive Agent (EEEC_Agent), which is capable of rear-end collision avoidance using artificial human driver emotions. However, the architecture of the EEEC_Agent is based on an ultrasonic sensory system which follows three-state driving strategies without considering the neighbor vehicles types. To address these issues, in this paper we propose an extended version of the EEEC_Agent which contains human driver-inspired dynamic driving mode controls for autonomous vehicles. In addition, a novel end-to-end learning-based motion planner has been devised to perceive the surrounding environment and regulate driving tasks accordingly. The real-time in-field experiments performed using a prototype AV demonstrate the effectiveness of this proposed rear-end collision avoidance system.

Human and Scene Motion Deblurring Using Pseudo-Blur Synthesizer

Present-day deep learning-based motion deblurring methods utilize the pair of synthetic blur and sharp data to regress any particular framework. This task is designed for directly translating a blurry image input into its restored version as output. The aforementioned approach relies heavily on the quality of the synthetic blurry data, which are only available before the training stage. Handling this issue by providing a large amount of data is expensive for common usage. We answer this challenge by providing an on-the-fly blurry data augmenter that can be run during training and test stages. To fully utilize it, we incorporate an unorthodox scheme of deblurring framework that employs the sequence of blur-deblur-reblur-deblur steps. The reblur step is assisted by a reblurring module (synthesizer) that provides the reblurred version (pseudo-blur) of its sharp or deblurred counterpart. The proposed module is also equipped with hand-crafted prior extracted using the state-of-the-art human body statistical model. This prior is employed to map human and non-human regions during adversarial learning to fully perceive the characteristics of human-articulated and scene motion blurs. By engaging this approach, our deblurring module becomes adaptive and achieves superior outcomes compared to recent state-of-the-art deblurring algorithms.

Learning Instance Motion Segmentation With Geometric Embedding

Most existing deep learning-based motion segmentation methods treat motion segmentation as a binary segmentation problem, which is generally not the real case in dynamic scenes. In addition, the object and camera motion are often mixed, making the motion segmentation problem difficult. This paper proposes a joint learning method which fuses semantic features and motion clues using CNNs with deformable convolution and a motion embedding module, to address multi-object motion segmentation problem. The deformable convolution module serves to fusion color and motion information. And the motion embedding module learns to distinguish objects' motion status with inspiration from geometric modeling methods. We perform extensive quantitative and qualitative experiments on benchmark datasets. Especially, we label over 9000 images of KITTI visual odometry dataset to help training the deformable module. Our method achieves superior performance in comparison to the current state-of-the-art in terms of speed and accuracy.

Conditional DQN-Based Motion Planning With Fuzzy Logic for Autonomous Driving

Motion planning is one of the most significant part in autonomous driving. Learning-based motion planning methods attract many researchers’ attention due to the abilities of learning from the environment and directly making decisions from the perception. The deep Q-network, as a popular reinforcement learning method, has achieved great progress in autonomous driving, but these methods seldom use the global path information to handle the issue of directional planning such as making a turning at an intersection since the agent usually learns driving strategies only by the designed reward function, which is difficult to adapt to the driving scenarios of urban roads. Moreover, different motion commands such as the steering wheel and accelerator are associated with each other from classic Q-networks, which easily leads to an unstable prediction of the motion commands since they are independently controlled in a practical driving system. In this paper, a conditional deep Q-network for directional planning is proposed and applied in end-to-end autonomous driving, where the global path is used to guide the vehicle to drive from the origination to the destination. To handle the dependency of different motion commands in Q-networks, we take use of the idea of fuzzy control and develop a defuzzification method to improve the stability of predicting the values of different motion commands. We conduct comprehensive experiments in the CARLA simulator and compare our method with the state-of-the-art methods. Experimental results demonstrate the proposed method achieves better learning performance and driving stability performance than other methods.

Deep learning-based image reconstruction and motion estimation from undersampled radial k-space for real-time MRI-guided radiotherapy

To enable magnetic resonance imaging (MRI)-guided radiotherapy with real-time adaptation, motion must be quickly estimated with low latency. The motion estimate is used to adapt the radiation beam to the current anatomy, yielding a more conformal dose distribution. As the MR acquisition is the largest component of latency, deep learning (DL) may reduce the total latency by enabling much higher undersampling factors compared to conventional reconstruction and motion estimation methods. The benefit of DL on image reconstruction and motion estimation was investigated for obtaining accurate deformation vector fields (DVFs) with high temporal resolution and minimal latency.2D cine MRI acquired at 1.5 T from 135 abdominal cancer patients were retrospectively included in this study. Undersampled radial golden angle acquisitions were retrospectively simulated. DVFs were computed using different combinations of conventional- and DL-based methods for image reconstruction and motion estimation, allowing a comparison of four approaches to achieve real-time motion estimation. The four approaches were evaluated based on the end-point-error and root-mean-square error compared to a ground-truth optical flow estimate on fully-sampled images, the structural similarity (SSIM) after registration and time necessary to acquire k-space, reconstruct an image and estimate motion.The lowest DVF error and highest SSIM were obtained using conventional methods up to . For undersampling factors , the lowest DVF error and highest SSIM were obtained using conventional image reconstruction and DL-based motion estimation. We have found that, with this combination, accurate DVFs can be obtained up to with an average root-mean-square error up to 1 millimeter and an SSIM greater than 0.8 after registration, taking 60 milliseconds.High-quality 2D DVFs from highly undersampled k-space can be obtained with a high temporal resolution with conventional image reconstruction and a deep learning-based motion estimation approach for real-time adaptive MRI-guided radiotherapy.

Spatio-temporal deep learning methods for motion estimation using 4D OCT image data

PurposeLocalizing structures and estimating the motion of a specific target region are common problems for navigation during surgical interventions. Optical coherence tomography (OCT) is an imaging modality with a high spatial and temporal resolution that has been used for intraoperative imaging and also for motion estimation, for example, in the context of ophthalmic surgery or cochleostomy. Recently, motion estimation between a template and a moving OCT image has been studied with deep learning methods to overcome the shortcomings of conventional, feature-based methods.MethodsWe investigate whether using a temporal stream of OCT image volumes can improve deep learning-based motion estimation performance. For this purpose, we design and evaluate several 3D and 4D deep learning methods and we propose a new deep learning approach. Also, we propose a temporal regularization strategy at the model output.ResultsUsing a tissue dataset without additional markers, our deep learning methods using 4D data outperform previous approaches. The best performing 4D architecture achieves an correlation coefficient (aCC) of 98.58% compared to 85.0% of a previous 3D deep learning method. Also, our temporal regularization strategy at the output further improves 4D model performance to an aCC of 99.06%. In particular, our 4D method works well for larger motion and is robust toward image rotations and motion distortions.ConclusionsWe propose 4D spatio-temporal deep learning for OCT-based motion estimation. On a tissue dataset, we find that using 4D information for the model input improves performance while maintaining reasonable inference times. Our regularization strategy demonstrates that additional temporal information is also beneficial at the model output.

Reinforcement Learning-Based Motion Planning for Automatic Parking System

In automatic parking motion planning, multi-objective optimization including safety, comfort, parking efficiency, and final parking performance should be considered. Most of the current research relies on the parking data from expert drivers or prior knowledge of humans. However, it is challenging to obtain a large amount of high-quality expert drivers’ data. Furthermore, expert drivers’ data or prior knowledge of humans does not guarantee an optimal multi-objective parking performance. In this article, we propose a model-based reinforcement learning method that learns parking policy of the data, by executing the data generation, data evaluation, and training network, iteratively. The trained network is used to guide the data generation cycle in the subsequent iteration. Based on this proposed method, we can get rid of human experience largely and learn parking strategies autonomously and quickly. The learned strategies ensure the multi-objective optimality of above requirements in the parking process. First, an environment model that approximates the actual environment is established, and the learning efficiency is accelerated through the simulated interaction between the agent and the environment model. To make the system independent of expert data or prior knowledge, a data generation algorithm combining Monte Carlo Tree Search (MCTS) and longitudinal and lateral policies is proposed. Then, to meet the multi-objective optimal demands mentioned above, a reward function is constructed to evaluate and filter the parking data. Finally, a neural network is used to learn the parking strategy from the filtered data. From the real vehicle test benchmarked with a mass-produced parking system, the proposed method is found to achieve better parking efficiency and lower requirements for start parking posture, thereby verifying the algorithm’s superiority.

Motion Prediction and Pre-Rendering at the Edge to Enable Ultra-Low Latency Mobile 6DoF Experiences

As virtual reality (VR) applications become popular, the desire to enable high-quality, lightweight, and mobile VR can potentially be achieved by performing the VR rendering and encoding computations at the edge and streaming the rendered video to the VR glasses. However, if the rendering has to be performed after the edge gets to know of the user’s new head and body position, the ultra-low latency requirements of VR will not be met by the roundtrip delay. In this article, we introduce edge intelligence, wherein the edge can predict, pre-render and cache the VR video in advance, to be streamed to the user VR glasses as soon as needed. The edge-based predictive pre-rendering approach can address the challenging six Degrees of Freedom (6DoF) VR content. Compared to 360-degree videos and 3DoF (head motion only) VR, 6DoF VR supports both head and body motion, thus not only viewing direction but also viewing position can change. Hence, our proposed VR edge intelligence comprises of predicting both the head and body motions of a user accurately using past head and body motion traces. In this article, we develop a multi-task long short-term memory (LSTM) model for body motion prediction and a multi-layer perceptron (MLP) model for head motion prediction. We implement the deep learning-based motion prediction models and validate their accuracy and effectiveness using a dataset of over 840,000 samples for head and body motion.

Phase-Based Noncontact Vibration Measurement of High-Speed Magnetically Suspended Rotor

Video-based measurement, as one of the most common noncontact measurement methods for industrial inspection, has been developed rapidly. Digital video cameras offer the benefit of low cost, high automation, and are capable to make the simultaneous full-field measurement. This article focuses on studying the microvibrational signal extraction for high-speed rotating machinery with digital cameras. An improved phase-based motion extraction and learning-based video magnification are proposed for measuring the microvibration with high frequency and small amplitude. The phase-based motion extraction is improved by directly transforming the phase variations into the displacement without the computation of phase gradient. Furthermore, the learning-based motion magnification is utilized to amplify and qualitatively measure the microvibration in specified frequency bands. The phase-based motion extraction is a quantitative measurement, while the learning-based video magnification is a qualitative measurement. Experimental results for microvibration measurement rising from a magnetically suspended motor system validate the effectiveness of the proposed method.

Attention-aware fully convolutional neural network with convolutional long short-term memory network for ultrasound-based motion tracking.

One of the promising options for motion management in radiation therapy (RT) is the use of LINAC-compatible robotic-arm-mounted ultrasound imaging system due to its high soft tissue contrast, real-time capability, absence of ionizing radiation, and low cost. The purpose of this work is to develop a novel deep learning-based real-time motion tracking strategy for ultrasound image-guided RT. The proposed tracker combined the attention-aware fully convolutional neural network (FCNN) and the convolutional long short-term memory network (CLSTM) that is end-to-end trainable. The glimpse sensor module was built inside the attention-aware FCNN to discard majority of background by focusing on a region containing the object of interest. FCNN extracted discriminating spatial features of glimpse to facilitate temporal modeling for CLSTM. The saliency mask computed from CLSTM refined the features particular to the tracked landmarks. Moreover, the multitask loss strategy including bounding box loss, localization loss, saliency loss, and adaptive loss weighting term was utilized to facilitate training convergence and avoid over/underfitting. The tracker was tested on the databases provided by MICCAI 2015 challenges, and the ground truth data were obtained with the help of brute force-based template matching technology. The mean tracking error of 0.97±0.52mm and maximum tracking error of 1.94mm were observed for 85 point landmarks across 39 ultrasound cases compared to the ground truth annotations. The tracking speed per frame per landmark with the GPU implementation ranged from 66 and 101 frames per second, which largely exceeded the ultrasound imaging rate. The results demonstrated the robustness and accuracy of the proposed deep learning-based motion estimation, despite the existence of some known shortcomings of ultrasound imaging such as speckle noise. The tracking speed of the system was found to be remarkable, sufficiently fast for real-time applications in RT environment. The approach provides a valuable tool to guide RT treatment with beam gating or multileaf collimator (MLC) tracking in real time.

Learning Human–Robot Interaction for Robot-Assisted Pedestrian Flow Optimization

Due to the fast-is-slower phenomenon in emergency escape, it is desirable to regulate pedestrian flows at the exit or a bottleneck. We propose a new robot-assisted pedestrian regulation and study passive human–robot interaction (HRI). A learning-based motion control approach is presented for a robot to efficiently interact with pedestrians for desirable collective motion. We first formulate the problem into an optimal control framework using the pedestrian dynamics description based on existing social force models with embedded HRI forces. To solve the defined optimal control problem, we propose an adaptive dynamic programming (ADP) approach to provide adjustable motion parameters of the robot to efficiently interact with pedestrians so that the regulated pedestrian flow tracks a desired velocity. The ADP control process only uses observed flow information rather than the models of pedestrians, and the ADP method provides feedback control with online learning and control capability. Simulation results demonstrate that the proposed approach can regulate pedestrian flows to desirable speeds by online learning.