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.
Introduced by Mate Chokeria of Columbia University, Liu argues that collecting enough real robot demonstrations may be impractical: one estimate puts the equivalent of human-level robot experience at 100,000 years. She proposes using existing models as teachers that generate demonstrations, rather than expecting each model to act as a general-purpose policy.
Liu describes VLK, a pipeline that combines 3D Gaussian splatting reconstructions with synthesized, trackable robot motion to generate vision, language, and kinematics supervision. The team generated 48,000 trajectories in a few hours and transferred a policy trained on synthetic data directly to real-world navigation and whole-body manipulation. A contact-conditioned motion-tracking policy also supports tasks such as sitting, stepping over obstacles, and handling objects.
For articulated tools such as tongs and syringes, small changes in contact and finger position can disrupt delicate force balance, making demonstrations and long-horizon reinforcement learning difficult. Liu’s team generates batches of short motion clips using grasp generation, reinforcement learning, motion planning, and interpolation, then trains a policy to connect them. Specialized parkour policies provide trajectories for training a more general student, while hindsight physics helps train a humanoid to adjust its compliance without altering reference motions.
Liu presents a model-based method for transferring a pretrained gripper policy to unseen hands by masking the robot’s appearance in the vision encoder and mapping actions and proprioception between embodiments. The method transfers without dexterous-hand training data. She also describes play pretraining for precise assembly: a policy fine-tuned on play data learned submillimeter insertion, while training the assembly task from scratch did not succeed.
Liu connects robot learning to cognitive, associative, and autonomous stages of skill development, and suggests progress may come from specialized models coordinated by capable orchestrators rather than one model that does everything. She argues models can earn a role when their assumptions and interfaces are clear. In the Q&A, she describes transferred policies as starting points that can be fine-tuned, and says better interfaces could help orchestrators combine models.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
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.
