Robotic dexterity with intrinsic sensing and reinforcement learning
This technology is a robot hand that uses model-free deep reinforcement learning to perform dexterous manipulation tasks.
Unmet Need: Robotic in-hand manipulation for dexterous tasks
Dexterous in-hand manipulation is a vital step towards achieving complex manipulation tasks through robotics, and yet it remains largely unsolved. While existing work shows some promise, these methods require large training times, cannot be used for arbitrary orientations, and require extensive external sensing involving multi-camera systems.
The Technology: Robotic platform utilizing model-free reinforcement learning, tactile sensing, and proprioception
This technology is a robotic hand that uses model-free deep reinforcement learning combined with tactile sensing and proprioception to achieve in-hand reorientation of objects. The robot can learn finger gating using only precision grasps, and can generalize these skills to objects not included in training. Only internal sensing is used, eliminating the need for external cameras; as such the robot is robust to visual distractions and perturbations. The technology enables continuous object reorientation around a specified axis, and can perform more complicated tasks than pick-and-place robots, for a variety of objects
Applications:
- eCommerce and logistics
- Manufacturing
- Remote maintenance
- Waste sorting
- General dexterous robotic control of an end-effector
- Translation and rotation of generic objects
- Robotic manipulation in a visually distracting environment
Advantages:
- Utilizes finger-gaiting and in-grasp manipulation for possible rotation of objects
- Capable of more complicated tasks than pick-and-place tasks
- Robust to changes in illumination and other visual distractions
Lead Inventor:
Patent Information:
Related Publications:
Tech Ventures Reference:
IR CU22008
Licensing Contact: Greg Maskel