RI Seminar: Sangbae Kim : Physical Intelligence and Cognitive Biases Toward AI

In this talk

1. 3:22

   Why robots need physical intelligence

   Kim argues that robots designed for precise position control work well in factories but struggle in uncontrolled environments, where surfaces, loads, and contact vary. He describes bio-inspired work including Stickybot’s gecko-like directional adhesive and the MIT Cheetah, whose compliant actuation helped it balance and jump over obstacles.

2. 17:49

   Combine learned behavior with reasoning

   Kim contrasts data-driven methods such as reinforcement learning, which can cover many situations but may be hard to steer, with model-based methods such as MPC, which are easier to reason about but can fail when their assumptions break. He gives examples of combining them, including shaping reinforcement-learning costs with physical models and embedding optimization tools in neural networks.

3. 23:31

   Human judgments can misread AI

   Kim says people often judge robots by human social instincts, so a backflip can seem more impressive than the demanding work of balancing. He notes that viewers may interpret a robot being pushed as aggression, even though that reaction says more about human perception than the robot’s intelligence.

4. 33:14

   Everyday skills exceed our verbal descriptions

   Kim points to chewing, reaching into a pocket, and grasping fruit as skills people perform without knowing their own movements or force profiles. He argues that ordinary language such as “spread” leaves out the detailed physical information a robot needs, while human action can adapt through rapid, mostly unconscious adjustments.

5. 51:54

   Build grounded layers for robust robots

   Kim distinguishes digital tasks, where execution is predictable, from physical tasks, where contact and conditions introduce uncertainty. He proposes low-level modules that ground high-level plans, handle fast interactions, and can be checked or patched, alongside separate safety systems. He says this hierarchy is meant to ground abstract plans, not simply imitate biological reflexes.