On February 7, 2025, during the paper presentation at the Center for Scientific Innovation and Education, speaker Nane Hakhverdyan presented the latest advancements in robotics related to real-time control policies on legged robots using model-predictive path integral control (MPPI).

Building robots that can gracefully traverse difficult terrains and skillfully manipulate objects like humans and animals has been a long-standing goal in robotics. For a long time, the topics of robot locomotion and manipulation have been studied separately. Recent interest in generalpurpose robot agents (i.e. humanoid robots) in both industry and academia has motivated designing control and planning algorithms capable of mastering both locomotion and manipulation skills in the same embodiment.

The paper takes advantage of the increasing performance of modern-day robotics simulation technology, particularly the MuJoCo physics engine, to compute real-time control policies on legged robots using model-predictive path integral control (MPPI). Contrary to the common belief that sampling approaches in high-dimensional tasks are computationally intractable, especially in real-time scenarios, the findings indicate that MPPI can be surprisingly effective in solving legged robot locomotion and manipulation tasks with a few simple design choices. To the best knowledge of the authors, this work represents the first successful deployment of sample-based whole-body control on real-world legged robots.