Research

Learning multiple skills for adaptive legged locomotion

Multi-Expert Learning Architecture - MELA - which fuses multiple expert neural networks into a synthesized expert and learns versatile motor skills. MELA can generate adaptive motor behaviors across different locomotion modes and survive in unknown situations, which is essential for truly autonomous robots to operate effectively in the wild.

Learning to catch flying objects reactively

Learning to reach and grasp flying objects of different shapes (spherical, cylindrical and cubic objects)

Tactile sensing for fine grasping and manipulation

Anthropomorphic robotic hand for manipulation and grasping

Haptic tele-manipulation

Fairspace live demo and streaming of haptic-guided teleoperation: exploit the sense of interaction forces and visual feedback for space assembly.

Fine control of grasping force

Compliant force control for strawberry-picking.

Immersive virtual reality for manipulation

Immersive multimodal interface to handle and manipulate various objects: immersive VR headset, optical markerless hand-tracking and haptic gloves.

Autonomous robotic manipulation: reactive grasping

Learning reactive reaching, graspign and re-grasping to handle slippage and disturbances.

Autonomous fall recovery

Learning whole-body motor skills for legged robots. The learning algorithm can successfully discover various recovery behaviors by exploiting whole-body contacts with the environment to achieve dynamic and contact-rich tasks.

Learning obstacle negotiation and fall recovery for a delivery robot

Object detection and semantic segmentation

Visual perception for robot manipulation and grasping.

Robust walking and push recovery

Optimisation and predictive control to counterbalance external pushes.