[ICON Spring26 Seminar] Harry Asada (MIT) #robotics #control #mpc

In this talk

1. 22:02

   Robotics beyond data

   Asada argues that robotics has deeper questions than collecting data or applying reinforcement learning. He recalls early work at Kyoto University, including a model of grasp stability based on potential energy and a serial elastic actuator that measured force through spring deformation. He urges researchers to keep studying the body and the physical principles that can inform machines.

2. 26:46

   Robots that work alongside people

   Asada describes direct-drive robots as clean and backdrivable, making them useful for physical interaction and for recording demonstrations of human skill. He later developed extra robotic limbs for aircraft assembly and astronaut assistance, and discusses a Tai Chi support robot for older adults. With Stanford collaborators, his team is also building a flexible robot that slides beneath a person to help caregivers transfer them from a bed to a wheelchair.

3. 35:27

   Koopman operators for changing dynamics

   Asada presents Koopman operators as a way to represent nonlinear dynamics linearly in a higher-dimensional space, potentially covering multiple contact modes with one model. His group adapted the method for controlled systems to produce a linear model suitable for control. He pairs it with model predictive control, which can plan through changes such as a robot leg moving from flight to ground contact.

4. 54:59

   Applications in walking and object pushing

   For a six-spoke rimless wheel, Asada’s team models airborne and contact phases together and uses MPC to choose actions across them; longer prediction horizons made forward motion more reliable. In a separate puck-pushing demonstration with a UR robot, the controller reacts to disturbances without a preset contact sequence. Control calculations averaged 5.3 milliseconds, compared with nearly 100 times longer for the traditional mixed-integer nonlinear method.

5. 1:09:38

   Human movement as a research guide

   Asada says Koopman-based control offers a way to predict the effects of complex robot actions while speeding computation. He points to muscle synergies, which coordinate many muscles through a smaller set of variables, as a possible clue for controlling high-degree-of-freedom robots. His advice to students is to stay curious: much of the body’s complexity remains unexplored.