Research
My research focuses on the convergence of robotics and biology. By understanding and abstracting the design principles found in biological systems, I cultivate expertise in several key areas, including modelling and transfer of human/animal excellence/knowledge for autonomous motion learning, detection & generation, enhancement of sensory-motor and learning capabilities, optimized morphological design for behavioural variability, optimal dynamics control for motor control learning. My primary objectives revolve around enhancing the comprehension of the biological underpinnings for lifelong learning, including adaptability and autonomy exhibited by animals, as well as creating innovative machine intelligence solutions, robotic systems, and applied control systems characterized by greater adaptability, resilience, and energy efficiency.
My research efforts encompass various specific areas, including human-robot collaboration & interaction, agriculture & aquaculture environment control, soft robotics & new actuations, rehabilitation, and embodied artificial intelligence.
Motion/grasp planning and optimization in complex environments based on interactions and demonstration-based learning
Adaptation to unorganized, congested and uncertain environment is a desirable capability but challenging task in development of robotic motion planning algorithms for object grasping. We have to make a tradeoff between coping with the environmental complexities using computational expensive approaches, and enforcing practical manipulation and grasping in real-time.
Motion planning is an essential aspect for robotic system and control. Prevailing solutions such as sampling-based push-grasping, optimization-based approaches and learning from demonstration (LfD) have demonstrated their effectiveness in specific motion planning task. However, sampling-based approaches (e.g., RRT, PRM) are computationally inefficient for manipulator motion planning, optimization-based methods typically require hand-coded cost functions to achieve and formulate the desired behaviors. In unorganized, congested and complex environments diffused with rigid and deformable objects, robotic motion planning for object grasping has been proven an intractable task. LfD approaches guided by human expert are plausibly the more approachable solutions towards this problem, because of their implicit behavior learning from expert demonstrations and less dependence on analytical models. Nevertheless, cluttered environment brings some challenging issues need to be solved under this method, for instance, coping with deformable obstacles. Deformable objects (e.g., flexible pipes, clothes, curtains) present different features from rigid ones because of their texture and deformability, and their infinite degrees of freedom make the configuration very intractable to recover. They may also cause uncontrolled robotic motion during grasping and/or manipulation. Towards this end, the preliminary research questions that we are approaching include how to optimally grasp and rearrange the movable deformable obstacles to reach the target object? And how to find good grasp configurations that are reachable in a cluttered environment?
The 7-DOF state-of-the-art robot Franka Emika is used for this research. Inspired by human agility and sense of touch, it is a sensitive and extraordinarily versatile power tool. With torque sensors in all seven axes, the arm skillfully and delicately manipulates objects, flawlessly accomplishing tasks you program it for (check here).
This research is based on the idea of expert guided trajectory optimization by LfDs as a strong influence on robotic motion planning for grasping. In real-world cluttered scenarios, it might be impossible for the robot to directly reach the target by following a collision-free trajectory and traversing between static and movable objects. Thus, this work has potential applications in push-grasping in various domains and scenarios diffused with rigid and deformable objects. For example, automated vegetable harvesting and fruit picking in agriculture, and autonomous garbage (e.g., protection suit, respirator, slightly flexible wires, and curtains) sorting in nuclear facilities. We use Probabilistic Movement Primitives (ProMPs) to learn the trajectory distribution from which any number of trajectories could be sampled. The figure on the right shows the end-effector trajectories in task space after conditioning from the learned distribution.
Design and control of bio-inspired/soft robotics: self-propelled vibro-driven mobile robots, and rigid/flexible arm and gripper systems
Vibro-driven robotic (VDR) systems use stick-slip motions for locomotion. Due to the underactuated nature of the system, efficient design and control are still open problems. In this research, we present a new energy preserving design, aiming to control the motion and improve the energy efficiency of the VDR system by leveraging from the passive dynamics of a spring-augmented pendulum. We indirectly control the friction-induced stick-slip motions by exploiting the passive dynamics in order to achieve an improvement in overall travelling distance and energy efficacy. Both collocated and non-collocated constraint conditions are elaborately analysed and considered to obtain a desired trajectory generation profile. For tracking control, we develop a partial feedback controller which for the pendulum which counter-acts the dynamic contributions from the platform. The proposed self-propelled robot is to the best of our knowledge the first nonlinear-motion prototype in literature towards the VDR systems.
The robotic platform is propelled over a surface rectilinearly via the interaction between driving forces and the horizontal sliding friction, resulting into an alternative sticking and slipping locomotion. Meanwhile, the elastic potential energy is stored and released alternatingly and synchronized with the contraction and relaxation of the torsional spring. The motion of the platform starts in a static state, and the robot moves when the magnitude of the resultant force applied on robot’s body in the horizontal direction exceeds the frictional force. Our model is developed to exploit feasible control approaches towards friction and stick-slip vibration to generate a periodic locomotion where the platform and the driving pendulum synchronize their motions harmoniously. In the sticking phase, the magnitude of the resultant force applied on the robot’s body in horizontal direction is less than the maximal static friction force. In the slipping phase, the resulting force is larger than the maximal static friction force. When this condition is met, the switches from sticking phase to slipping phase and the robot starts to move.
In our model, the harmonic property is introduced by our impedance model. The system performance is mainly determined by the counterbalance between rising and falling edges of the harmonic force. For each motion cycle, the displacement obtained in the forward motion stage is partially counteracted by the following backward motion causing a sub-optimal energy efficiency of the system. Therefore, we develop a two-stage trajectory profile to achieve our desired objectives that uses harmonic ramping edges in the forward motion stage and sufficiently neutralizes backward motions triggered by the falling edges.
The trajectory is composed of the locomotion and restoring stages. In the locomotion stage, the pendulum is driven with high acceleration using the release of elastic energy stored in the torsional spring. The excitation frequency of the harmonic force is considered in this stage to synchronize the pendulum motion with the ramping edge. The resultant interaction force generates a slipping motion. In the restoring stage, the pendulum is carefully returned to the initial position, and potential energy is simultaneously stored in the torsional spring for the next motion cycle. The resultant interaction force in the horizontal direction is less than the maximum dry friction force, i.e., the robot remains in the sticking phase in this stage.
Prosthetics and rehablitations using robotics and deep learning through EEG/ sEMG signals and Brian-Computer Interfaces (BCI)
Prosthetics and rehablitations using robotics and deep learning through EEG/ sEMG signals and Brian-Computer Interfaces (BCI). Rehabilitation and healthcare using robotics and artificial intelligence are important and innovative research areas.
Effect of shoulder angle variation on sEMG-based elbow joint angle estimation
The estimation scheme of four proposed methods. Method One: using a training set including all shoulder angles' training data; Method two: adding two shoulder muscles's sEMG as additional inputs; Method three: a two-step method using arm muscles' sEMG and two shoulder muscles' sEMG; Method four: a two-step method using arm muscles' sEMG and measured shoulder angle value by a motion sensor.
Cyber Physical Systems, Internet of Things, and Human-in-the-Loop Shared Control
Applications of thematic and emerging technologies including teleoperation, multi-agent systems and cooperative control, effectively and efficiently managing, analysing and exploring Cyber-Physical Systems, Internet of Things, Human-in-the-Loop Shared Control, etc. into fields of robotics, agri-food, future manufacturing and healthcare.