Mingyi Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mingyi Liu has a total of 19 co-authors. Here are the top co-authors by number of shared publications:
- Amir Degani
Technion—Israel Institute of Technology
1 shared publication Daniel Q Tan
1 shared publication- Daniel Qi Tan
Department of Materials Science and Engineering, Guangdong Technion-Israel Institute of Technology, 241 Daxue Road, Shantou 515063, PR China
1 shared publication - Izhak Bucher
Dynamics Laboratory, Mechanical Engineering
1 shared publication Morel Groper
1 shared publication- Muxin Weng1 shared publication
Latest Publications
- Open Access
- 12 August 2026
- Published by Elsevier in SSRN Electronic Journal
Energy-Efficient Bipedal Foot Placement via Torque-Sign-Constrained Expert Routing
Energy-efficient bipedal walking on uneven terrian requires precise foot placement on constrained targets with limited actuation. In reinforcement learning framework, torque, power, or action penalties do not prevent positive and negative hip torques from alternating within one control transition, which may result in antagonistic actuation and hence inefficient control sequense. We propose two proximal policy optimization (PPO) experts with non-positive or non-negative values actuation in the hip and a select expert to ensure none-antagonistic action, leading to precise and energy-efficient foot placement. Performance is evaluated using normalized positive mechanical work, denoted by C_mt. Simulation validation with two-link human walking model show that the mean C_mt was 0.122 for the proposed method, compared with 0.227 for unrestricted discrete PPO, 0.238 for continuous-torque PPO, and 0.199 for continuous-torque model predictive control. Across additional randomized four-link batches comprising 10,800 matched trials, the proposed method and active PPO achieved non-differentiable success rates. Among the 2,244 success pairs for both controllers, mean C_mt was 0.447 for the proposed method and 1.53 for active PPO ("proposed" - "active" paired difference is -1.08; 95% CI ranges from -1.19 to -0.976); the proposed method produced lower value in 81.8% of pairs. Separately, a single-factor query-schedule comparison on 2,160 matched cases showed that holding the selected expert throughout each transition increased success rate by 3.29 percentage points relative to per-control-step requerying (95% CI, 2.41--4.17). Prototype trials demonstrate foot-placement feasibility. These results show a reproducible simulated C_mt advantage for sign-constrained expert routing and a success-rate benefit from transition-level persistence under frozen-policy evaluation.
- 12 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEnergy-Efficient Bipedal Foot Placement via Torque-Sign-Constrained Expert Routing
Energy-efficient bipedal walking on uneven terrian requires precise foot placement on constrained targets with limited actuation. In reinforcement learning framework, torque, power, or action penalties do not prevent positive and negative hip torques from alternating within one control transition, which may result in antagonistic actuation and hence inefficient control sequense. We propose two proximal policy optimization (PPO) experts with non-positive or non-negative values actuation in the hip and a select expert to ensure none-antagonistic action, leading to precise and energy-efficient foot placement. Performance is evaluated using normalized positive mechanical work, denoted by C_mt. Simulation validation with two-link human walking model show that the mean C_mt was 0.122 for the proposed method, compared with 0.227 for unrestricted discrete PPO, 0.238 for continuous-torque PPO, and 0.199 for continuous-torque model predictive control. Across additional randomized four-link batches comprising 10,800 matched trials, the proposed method and active PPO achieved non-differentiable success rates. Among the 2,244 success pairs for both controllers, mean C_mt was 0.447 for the proposed method and 1.53 for active PPO ("proposed" - "active" paired difference is -1.08; 95% CI ranges from -1.19 to -0.976); the proposed method produced lower value in 81.8% of pairs. Separately, a single-factor query-schedule comparison on 2,160 matched cases showed that holding the selected expert throughout each transition increased success rate by 3.29 percentage points relative to per-control-step requerying (95% CI, 2.41--4.17). Prototype trials demonstrate foot-placement feasibility. These results show a reproducible simulated C_mt advantage for sign-constrained expert routing and a success-rate benefit from transition-level persistence under frozen-policy evaluation.
- Open Access
- 28 July 2026
- Published by Elsevier in SSRN Electronic Journal
Emergent Predictive Target Interception for Autonomous Surface Vehicles Using Onboard Sensing
Moving-target interception by autonomous surface vehicles (ASVs) in confined, cluttered workspaces is challenging under nonlinear dynamics, actuation delay, and onboard-sensor partial observability. This paper presents a data-driven predictive interception framework. It combines a residual data-driven dynamics model (DDM), privileged training with proximal policy optimization (PPO), and teacher--student distillation. The DDM learns body-frame velocity transitions from motion-capture trajectories and action histories without calibrated hydrodynamic coefficients. The teacher uses privileged simulation states during training, whereas the deployed student uses only a history of range rays, camera-derived target indicators, and the previous action. The teacher reaches a final simulation success rate of $99.25\pm0.32\%$ and maintains $98.04\pm0.51\%$ success with a 4.8~s actuation delay. In 100 randomized comparisons, PPO achieves 97\% success versus 99\% for a DDM-based model predictive control (MPC) baseline, while reducing mean interception time from 62.2~s to 55.9~s and evaluation time from 440.9~s to 1.5~s. Trajectory and lead-angle analyses show that the learned policy steers toward future capture regions rather than following the instantaneous line of sight. The student network can distill the teach policy and achieve 80\% of success rate even under strigent "success" standards. Pool experiments further demonstrate interception using onboard sensing only.- 28 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEmergent Predictive Target Interception for Autonomous Surface Vehicles Using Onboard Sensing
Moving-target interception by autonomous surface vehicles (ASVs) in confined, cluttered workspaces is challenging under nonlinear dynamics, actuation delay, and onboard-sensor partial observability. This paper presents a data-driven predictive interception framework. It combines a residual data-driven dynamics model (DDM), privileged training with proximal policy optimization (PPO), and teacher--student distillation. The DDM learns body-frame velocity transitions from motion-capture trajectories and action histories without calibrated hydrodynamic coefficients. The teacher uses privileged simulation states during training, whereas the deployed student uses only a history of range rays, camera-derived target indicators, and the previous action. The teacher reaches a final simulation success rate of $99.25\pm0.32\%$ and maintains $98.04\pm0.51\%$ success with a 4.8~s actuation delay. In 100 randomized comparisons, PPO achieves 97\% success versus 99\% for a DDM-based model predictive control (MPC) baseline, while reducing mean interception time from 62.2~s to 55.9~s and evaluation time from 440.9~s to 1.5~s. Trajectory and lead-angle analyses show that the learned policy steers toward future capture regions rather than following the instantaneous line of sight. The student network can distill the teach policy and achieve 80\% of success rate even under strigent "success" standards. Pool experiments further demonstrate interception using onboard sensing only. - Open Access
- 23 June 2026
- Published by Elsevier in SSRN Electronic Journal
History-Augmented Data-Driven Dynamics Modeling with Residual-Gated Switching for Active Vibration Suppression
Active vibration suppression in mechatronic hardware is complicated by hidden actuation memory: the commanded actuator input does not fully determine the force that drives the structure because of actuator dynamics, delay, saturation, nonlinearities, and unmeasured input perturbations. This paper proposes a unified finite-history DDM--AR-PPO--DGS framework for this problem. A history-augmented data-driven dynamics model (DDM) is identified from hardware rollouts and used both as a surrogate environment for adaptive-range residual proximal policy optimization (AR-PPO) and as a nominal one-step predictor for dynamics-guided switching (DGS). The full DDM reduces held-out open-loop SW-NRMSE from 0.608±0.260 for a memoryless baseline to 0.138±0.045, i.e., a 77.3 \% reduction in error. In hardware tests under known shaker-side excitations, AR-PPO gives the lowest sweep-averaged payload velocity and acceleration RMS among the tested causal controllers, while displacement suppression remains competitive with the model-based baselines. Under unknown actuator-side input pulses, DGS-AR-PPO detects all tested pulses and reduces event-window displacement, velocity, and acceleration RMS by 31.1%, 43.5%, and 25.2%, respectively, relative to standalone AR-PPO. The conclusions are restricted to the tested two-degree-of-freedom, single-actuator platform and disturbance envelope.
- 23 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHistory-Augmented Data-Driven Dynamics Modeling with Residual-Gated Switching for Active Vibration Suppression
Active vibration suppression in mechatronic hardware is complicated by hidden actuation memory: the commanded actuator input does not fully determine the force that drives the structure because of actuator dynamics, delay, saturation, nonlinearities, and unmeasured input perturbations. This paper proposes a unified finite-history DDM--AR-PPO--DGS framework for this problem. A history-augmented data-driven dynamics model (DDM) is identified from hardware rollouts and used both as a surrogate environment for adaptive-range residual proximal policy optimization (AR-PPO) and as a nominal one-step predictor for dynamics-guided switching (DGS). The full DDM reduces held-out open-loop SW-NRMSE from 0.608±0.260 for a memoryless baseline to 0.138±0.045, i.e., a 77.3 \% reduction in error. In hardware tests under known shaker-side excitations, AR-PPO gives the lowest sweep-averaged payload velocity and acceleration RMS among the tested causal controllers, while displacement suppression remains competitive with the model-based baselines. Under unknown actuator-side input pulses, DGS-AR-PPO detects all tested pulses and reduces event-window displacement, velocity, and acceleration RMS by 31.1%, 43.5%, and 25.2%, respectively, relative to standalone AR-PPO. The conclusions are restricted to the tested two-degree-of-freedom, single-actuator platform and disturbance envelope.
- 27 December 2025
- Published by Elsevier in Ocean Engineering
- Vol. 348
A data-driven dynamic modeling of fish-like robot enabled by artificial neural network
- 27 December 2025
- Published by Elsevier in Ocean Engineering
- Vol. 348
A data-driven dynamic modeling of fish-like robot enabled by artificial neural network
- 26 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 10 (4) , 3996-4003
Improving the Energy Efficiency by Using Quasi-Passive-Dynamics-Based Elastic Actuator
Electric actuators are commonly used in machines and the optimal energy efficiency can be achieved only at certain speeds. This letter introduces the Quasi-Passive-Dynamics-based Elastic Actuator (QPD-EA), which enhances energy efficiency by leveraging passive dynamics and energy recycling. Instead of directly linking the motor to the robot arm, the motor charges a spring at its optimal energy efficiency speed, and a clutch mechanism controls energy transfer between the spring and the robot arm, enabling efficient movement in any desired direction. The design, modeling, prototype, and experimental validation are described in detail, showing that the QPD-EA consumes 86% less energy than conventional actuators for the same motion, reducing energy use from 154 mJ to 21 mJ. This approach can significantly improve the energy efficiency of actuators in pick-and-place operations.
- 26 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 10 (4) , 3996-4003
Improving the Energy Efficiency by Using Quasi-Passive-Dynamics-Based Elastic Actuator
Electric actuators are commonly used in machines and the optimal energy efficiency can be achieved only at certain speeds. This letter introduces the Quasi-Passive-Dynamics-based Elastic Actuator (QPD-EA), which enhances energy efficiency by leveraging passive dynamics and energy recycling. Instead of directly linking the motor to the robot arm, the motor charges a spring at its optimal energy efficiency speed, and a clutch mechanism controls energy transfer between the spring and the robot arm, enabling efficient movement in any desired direction. The design, modeling, prototype, and experimental validation are described in detail, showing that the QPD-EA consumes 86% less energy than conventional actuators for the same motion, reducing energy use from 154 mJ to 21 mJ. This approach can significantly improve the energy efficiency of actuators in pick-and-place operations.