The Movement Lab
Welcome to The Movement Lab!
The goal of our lab is to create coordinated, functional, and efficient whole-body movements for digital agents and for real robots to interact with the world. We focus on holistic motor behaviors that involve fusing multiple modalities of perception to produce intelligent and natural movements. Our lab is unique in that we study “motion intelligence” in the context of complex ecological environments, involving both high-level decision making and low-level physical execution. We developed computational approaches to modeling realistic human movements for Computer Graphics and Biomechanics applications, learning complex control policies for humanoids and assistive robots, and advancing fundamental numerical simulation and optimal control algorithms. The Movement Lab is directed by Professor Karen Liu.
Generalist Humanoids with Embodied Intelligence
In the pursuit of an embodiment of human intelligence, capable of performing human tasks in the human world, we believe that the "human body" is the most natural solution. To realize this vision, we develop humanoid policies that truly generalize to real-world human environments. Given the morphological similarities between humans and humanoids, we take the approach to train humanoid policies directly using human motion. This enables fluid, coordinated whole-body navigation, locomotion, and manipulation, reflecting both human physical and cognitive intelligence. Our lab explores various commercially available humanoid platforms and has developed an accessible, self-built mini-humanoid, Stanford ToddlerBot, to further demonstrate and refine embodied intelligence.
Learnable Physics Simulators for Humans, Robots and the World
Physics simulation is increasingly relied upon to predict the outcome of real-world phenomena. The rise of deep learning further augments the importance of physics simulation for training intelligent robots and embodied AI agents in safe and accelerated simulated environments. Our lab has created a number of physics simulation tools and algorithms that leverage both differential equations and measured data for building accurate simulation models of humans, robots, and the world they interact with.
Personalized Predictive Human Models
Developing personalized predictive human models to simulate human movements has a wide range of applications from accelerating sport medicine to designing exoskeletons to creating VR avatars. We aim to develop generalizable and computationally tractable models of natural motion with minimal amount of engineering effort. Our predictive human models can simulate a wide range of scenarios, while providing the option to be personalized to specific individuals using only a moderate amount of data. A key focus of our research is exoskeleton design. These models can be used offline to optimize exoskeleton parameters and online to estimate human kinematic and kinetic states using limited onboard sensors.
Dexterous Manipulation in Ecological Environments
Dexterous manipulation is a hallmark feature that enables humans to carry out day-to-day tasks in complex ecological environments, such as using a screwdriver, opening a water bottle, or folding bed sheets. To highlight the advantages of humanlike hands, we are interested in manipulation tasks that can only be achieved by utilizing multiple contacts, both intrinsic and extrinsic, with real-world objects, both near-rigid or deformable. We develop control policies utilizing non-trivial contact strategies for embodied AI agents and real robots.