Karen Liu, an associate professor of computer science at Stanford University, discusses physics-based human motion simulation and its use in robotics at an MIT Robotics talk.
Liu contrasts photorealistic digital humans with realistic behavior: the Siren demo looks convincing, but a live actress supplies its movement through motion capture. She argues that virtual humans should generate their own movements in response to goals and changes in their environment.
Liu’s lab uses anatomical models to make natural motion emerge from tasks and physics rather than copying reference trajectories. The team learns state-dependent joint torque limits and effort functions from OpenSim muscle simulations, then uses them in joint-level control. In examples such as jumping, conventional torque bounds let an agent adopt positions a real person could not use to generate force.
For reinforcement learning, the learned effort function shapes the reward, while learned torque limits constrain the agent’s actions. Liu also describes learned joint limits based on captured human poses; these make passive ragdoll motion more human-like and can prevent implausible poses and self-collisions in inverse kinematics.
Liu presents universal policies conditioned on dynamic model parameters, then adapts them using real-world trials. Rather than requiring the true physical parameters, the method searches for parameters that make the policy perform well. On a Darwin OP2 robot, about 20 walking sequences over 15 to 20 minutes improved forward motion, with some backward and sideways movement; later meta-learning work aimed to speed adaptation.
Liu describes simulated human partners for robot-assisted dressing, including models with weaker arms, tremors, or limited range of motion. Cloth simulation also provides training data to estimate pressure on limbs from the robot’s forces and torques. For exoskeleton fall prevention, virtual walking agents help train fall detectors and recovery policies, though the work shown remains in simulation.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Professor Karen Liu of Stanford gives the RSS 2026 closing keynote on using models to generate robot training data and improve policies without relying on ever larger real-world datasets.
Michael Chang, a PhD researcher at the University of California, Berkeley, presents his work at the GRASP Lab Spring 2023 SFI on neural abstractions for modeling and controlling systems, and discusses future human-computer interfaces at DeepMind.
Dennis Hong, robotics professor and founding director at Virginia Tech, explains why humanoid robots suit human environments and how his lab developed soccer-playing robots and firefighting prototypes.
A ROBOTIS presenter discusses a Darwin humanoid robot that plays Dance Dance Revolution and the vision-processing and motion constraints behind the demonstration.
Alex, role and affiliation unknown, presents the Darwin Drives project with King at an ICRA 2012 challenge; they show a humanoid robot operating an iRobot Create and discuss possible uses for humanoid-driven vehicles.
Students from Purdue University describe their team’s Darwin-OP robotics project at an ICRA 2012 challenge, explaining its sensor-based reactions and demonstrating them.
