Super-aligned Machine Intelligence via a Soft Touch

Physical embodiment is central in aligning human intentions with machine intelligence, where the moment of touch holds the truth between the simulated and real worlds. In this talk, I will argue for the foundational role of Vision-Based Tactile Sensing (VBTS) in developing robot learning systems toward embodied intelligence. Tactile sensing by vision provides modern robotic systems with shareable and reproducible access to a rich modality of real-time sensory representations in a multi-dimensional, computationally compatible data structure at a low cost. Drawing from our recent work in proprioceptive robot learning, I will explain our recipe in VBTS that combines soft robotic design with modern learning algorithms to achieve State-of-the-Art (SOTA) performances in tactile perception with omni-directional adaptation. Our VBTS solution is highly robust even after 1 million test cycles. Based on this principle, I will present how we achieve multi-modal outputs from a single, vision-based input by leveraging computational intelligence via various learning algorithms. Such unique capability enables us to incorporate tactile sensing with omni-directional adaptation in cross-limb, cross-skill, and cross-scenario applications, such as amphibious tactile grasping, proprioceptive and interoceptive shape reconstruction, industrial welding, and object detection. Then, I will present our preliminary results in developing a topological learning framework that simultaneously trains manipulation and locomotion skills by expanding tactile representation via a Graph Neural Network. Finally, I will present the portable DeepClaw system that captures human demonstration actions, including both force and motion, within a single device towards a universal interaction interface for embodied intelligence, the path towards super-aligned machine intelligence via a soft touch.

Post Talk Link:  Click Here

Passcode: +7AQ?G%k