RoboBoom
MIT Robotics· April 26, 2022· 49:30

MIT Robotics - Marco Hutter - Robots in the wild

Marco Hutter, professor at ETH and founder of ANYbotics, discusses legged robot control, autonomy, and field applications at an MIT Robotics talk.

In this talk

  1. 1:18

    Building robots for harsh environments

    Hutter describes research spanning robot design, model predictive control, reinforcement learning, autonomy, and perception. His team develops legged robots for settings such as mines, offshore sites, and search and rescue, while also applying robotics to construction, agriculture, and river cleanup.

  2. 4:23

    Combining control and perception

    Hutter explains how navigation, motion planning, whole-body control, and joint control work together, with feedback from both the robot’s own sensors and cameras or LiDAR. His team’s model predictive control uses terrain maps, contact constraints, collision avoidance, and limits on speed, torque, and friction to plan movement online.

  3. 11:34

    Learning to adapt to terrain

    Real terrain can be slippery, soft, or otherwise unlike the assumptions built into a controller. Hutter’s team trains policies in simulation, then uses robot experiments to learn actuator behavior and privileged training to help the robot infer terrain properties such as friction. In tests, the robot adapted after slipping on a soapy surface and could cross snow, grass, and difficult mountain terrain.

  4. 30:40

    Autonomous exploration and navigation

    For the DARPA Subterranean Challenge, Hutter’s team combined legged robots, drones, and wheeled vehicles, then used legged systems in the final. Robots explored tunnels, urban spaces, and caves, built maps, and searched for artifacts despite degraded perception and no communication. Four robots mapped several kilometers in an hour; a human mapping effort took 100 hours.

  5. 41:41

    From inspection to construction

    ANYbotics deploys quadrupeds for industrial inspection, using visual, thermal, and acoustic sensors to check gauges, detect temperature changes, and locate possible gas leaks. Hutter also describes learning data-driven control for a 12-ton hydraulic machine, enabling precise grading and autonomous construction. One project built a 10-meter-high, 80-meter-long stone structure using roughly 1,000 stones over a three-month construction site.

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

More talks