RoboBoom

Majid Khadiv: Control of legged robots in challenging environments

Majid Khadiv, a postdoc at the Max Planck Institute in Tübingen, discusses model-based control and machine learning for legged robots at a talk hosted by the Learning and Intelligent Systems Lab at TU Berlin.

In this talk

  1. 1:20

    Why legged robots are hard to control

    Khadiv points to hazardous work, assistive devices, and home care as potential uses for robots that can move with humanlike dexterity. Their control is difficult because contact changes abruptly, the robots can fall, and their movement is affected by disturbances and uncertainty. He combines model-based methods with machine learning to address these challenges.

  2. 5:27

    Adapting footsteps for robust walking

    Khadiv compares robots with fully actuated ankles, lightweight legs, and unactuated ankles, each with different limits on contact control. His controller adjusts the next footstep’s location and timing using center-of-mass feedback. On the Bolt robot, it handled uneven and soft terrain, pushes, and a moving platform that caused the feet to slip.

  3. 14:31

    Planning dynamic motions with full-body models

    For motions such as jumping and climbing, Khadiv uses centroidal momentum dynamics to plan center-of-mass movement, momentum, and contact forces. He iterates between that planner and whole-body solvers that enforce joint and torque limits. On the torque-controlled Solo robot, offline plans enabled trotting and a jump exceeding one meter, more than five times its 21-centimeter standing height.

  4. 21:02

    Using experimental data to tune robustness

    Khadiv describes using Bayesian optimization to tune the cost weights in a walking controller, balancing velocity tracking against contact stability and step margins. In simulation with actuator uncertainty, delays, and random pushes, the method found a controller that tracked the commanded speed and avoided falls after 13 optimization queries.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.