Workshop Program

Abstract:

Learning reliable and dexterous manipulation requires robots to generalize from limited interaction while accounting for uncertainty in their perception of the world. Geometry and embodiment provide strong inductive biases for this learning process, defining relationships that representations and policies should respect.

In this talk, I will present recent work on incorporating geometric structure into robot learning to enable more efficient, transferable, and reliable manipulation skills. I will demonstrate how morphological symmetry can be embedded into reinforcement learning and flow-based policies for dexterous and bimanual manipulation, allowing experience to be shared and transferred across symmetry-related configurations. I will then discuss uncertainty-aware representations that capture distributions over object poses and scene properties, supporting manipulation under occlusion, ambiguity, and partial observability.

Through these examples, I will argue that geometric structure shapes both how robots acquire skills and how they use those skills when their observations are uncertain. I will close with an outlook on how symmetry and uncertainty can jointly inform the development of robot learning systems that generalize reliably in the physical world.

Authors: Joshua R. Bettles, Yang Song, Bruno Vilhena Adorno, Joaquin Carrasco

Institution: The University of Manchester

Authors: Lorena Maria Genua, Sarvesh Prajapati, Ananya Trivedi, Taşkın Padır

Institution: Northeastern University

Authors: Anastasios Manganaris, Rohit Banerjee, Gokul Prabhakaran, Ahmed H. Qureshi

Institution: Purdue University

Authors: Zhefeng Huang, Yilin Cai, Yifan Wang, Yue Chen

Institutions: Georgia Institute of Technology; Emory University

Abstract:

Much of geometric robotics builds structure into the representation. This talk is about structure the robot already has: the metric its kinetic energy defines on configuration space, and the invariant submanifolds its dynamics carve out of the tangent bundle. Natural motion manifolds compactly represent the motions a system performs at low actuation cost, obtained from the conservative dynamics alone, without data. On Mody, a direct-drive double pendulum with a small fraction of the torque it would need to lift itself, this yields large-amplitude gymnastic swings on hardware. Preliminary results extend the idea to cyclic pick-and-place: for light payloads, the task becomes a periodic orbit whose turning points sit at the pick and place locations, leaving actuation to cover only dissipation. An outlook addresses heavier objects, where grasping alters the dynamics itself.

Abstract:

How can robots learn to predict and act in the complex physical world? In this talk, I will present Generative Particle World Models (GPWM), a learned physics engine that models 3D dynamics across rigid, deformable, cloth, and granular materials. GPWM can imagine physical futures far beyond its training horizon, generalize to novel objects and physical properties, and turn desired outcomes into robot actions through its learned dynamics. Together, these capabilities point toward generative 3D world models that can predict, imagine, and act in the physical world.

Authors: Chen Chen, Yunwen Li, Yifan Xu, Xiangjie Yan, Chang Shu, Jianxia Hou, Shiji Song, Xiang Li

Institutions: Tsinghua University; Tsingscribe Medical Ltd.; ETH Zurich; Peking University School and Hospital of Stomatology

Authors: Phone Thiha Kyaw, Jonathan Kelly

Institution: University of Toronto Institute for Aerospace Studies (UTIAS)

Authors: Riley Link, Pieter van Goor, Jake Welde

Institutions: Cornell University; University of Sydney

Abstract:

General-purpose manipulation requires robots to reason about structure at multiple levels, from continuous motion and physical interaction to uncertainty and long-horizon task composition. This talk presents recent work from the ELPIS Lab exploring these representations using contact-rich manipulation as a running example. It begins with geometric representations for generating task-space motions on redundant manipulators, then examines learned models of physical interaction and geometric reasoning about uncertainty during planning. It next considers how planning and execution can be coupled to adapt when reality deviates from prediction. Finally, the talk moves from continuous interaction to grounded symbolic representations that support long-horizon task planning. Together, these works illustrate how structured representations can connect motion, learning, planning, and abstraction toward more general-purpose robotic manipulation.