Cosimo Della Santina - Moving Without Thinking: Embodied Intelligence in Locomotion

In this talk

1. 1:13

   Intelligence from body mechanics

   Della Santina describes a dead fish swimming against a current, showing how the interaction between body and water can produce adaptive motion without a brain. He connects this idea to a robotic hand with two motors and 19 degrees of freedom, whose movements emerge from its mechanics and contact with objects.

2. 3:27

   Passive feet and legged bodies

   A passive robotic foot modeled on the human windlass mechanism adapts to uneven ground and smooths interaction forces. Its broad support region and shifting internal forces help it balance without active control. Della Santina also shows a motorless dog-like robot whose running motion is driven by a treadmill.

3. 7:56

   Using natural oscillations for locomotion

   For steady running, Della Santina’s group studies the natural oscillations of a robot’s joints and builds controllers to excite and sustain them. A spring-loaded segmented leg maintains oscillations at different energy levels despite disturbances. Combining these oscillations enabled whole-body systems to run or hop efficiently.

4. 10:29

   Planning jumps and combining intelligence

   For jumps, the group uses reduced-order models and numerical optimization to account for a robot’s elasticity, then retargets the result to whole-body control. A modified Go1 robot with engageable parallel springs lets them compare jumping with and without added elasticity; reinforcement learning supports more aggressive, precise jumps. Della Santina closes with work on co-learning a leg’s mechanics and control, and exploring locomotion with continuum soft robots.

5. 14:10

   Questions on foot elasticity and modeling

   Asked how elastic feet affect rollover and forward progression, Della Santina says the question remains untested in their system. He suggests the human foot’s windlass mechanism may store energy in the plantar fascia and help power a step. He explains that the dog motion was derived from movement data, reduced computationally, then retargeted using models of the robot.