MIT Robotics - Quan Nguyen - Hierarchical Learning and Control for Agile and Adaptive Legged Robots

In this talk

1. 6:24

   MPC for locomotion and manipulation

   Nguyen describes model predictive control for quadruped and humanoid robots, including a Cheetah 3 climbing stairs without cameras or prior stair information. He explains how hierarchical MPC lets a robot plan contact forces and locations to push a bulky box while controlling its position and orientation. For humanoid manipulation, he models a held object as an external force so picking it up or releasing it makes smaller changes to the controller.

2. 24:45

   Adapting to unknown loads and objects

   Nguyen presents an adaptive controller that compares a robot carrying an unknown load with a reference model, then adjusts control to compensate. The robot carried five kilograms, ran on a slope, and stayed stable while running with water onboard. A hierarchical extension estimates how to move an object with unknown mass or center of mass, then distributes forces and contact locations across the robot.

3. 31:45

   Coordinating robots on shared tasks

   Nguyen extends adaptive local manipulation to large objects and multiple robots. In one example, two robots carry an object over uneven terrain and coordinate their foot placement and body height to clear obstacles. He notes that perception and coordination become important when terrain includes gaps and stepping stones.

4. 38:30

   Reinforcement learning for agility

   Nguyen describes reinforcement learning that combines stance force control with swing-foot control, which are separate in the basic MPC setup. He reports up to 58% faster turning and load carrying up to 80% of a robot's weight on compliant, unstructured terrain. A policy trained in simulation on an A1 transferred to A1, Go1, and the larger AlienGo; another approach learns smooth transitions among walking, hopping, and jumping modes.