🤖🍨 Sundae Robotics 02: Robotics, Dexterity, Cross-Embodiment & World Models — DexterityGen & SPIDER

​🤖🍨 Grab a sundae and join Sundae Robotics, a private, invite-only Sunday series bringing together robotics researchers, founders, and builders working at the frontier of physical intelligence.

​Sundae Robotics 02
Human Demonstrations & Foundation Controllers
Featured Talk: DexterityGen & SPIDER

​Keynote: Changhao Wang
Postdoctoral Researcher, Stanford University (Shuran Song) · Former Meta FAIR · PhD, UC Berkeley (Masayoshi Tomizuka)

​Human dexterity is one of the richest and most scalable sources of data for robot learning. Even a seemingly simple task such as rotating a ball can be accomplished through many different strategies: coordinated finger motions, wrist rotations, or repeated regrasps. While robots may ultimately develop their own solutions, human demonstrations provide an abundant and expressive prior for learning general-purpose manipulation. The challenge is enabling robots with fundamentally different embodiments to interpret, retarget, and execute those demonstrations effectively.

​In this talk, Changhao will present two complementary directions toward bridging human dexterity and robotic control. First, DexterityGen introduces a foundation controller capable of producing an unprecedented range of dexterous manipulation behaviors across diverse tasks. Second, SPIDER (Scalable Physics-Informed DExterous Retargeting) presents a scalable, physics-informed retargeting framework that transfers human demonstrations to robot hands while respecting embodiment-specific kinematic and physical constraints. Together, these approaches demonstrate how robots can learn increasingly human-like manipulation from both teleoperation and large-scale collections of human videos, significantly expanding the data available for training general-purpose manipulation systems.

​Changhao's research spans dexterous manipulation, robot learning, teleoperation, sim-to-real transfer, tactile sensing, and contact-rich manipulation. His recent work includes DexterityGen (RSS 2025), Geometric Retargeting (IROS 2025), Dexterity from Smart Lenses (ICRA 2026), DexCtrl (ICRA 2026), OSMO: Open-Source Tactile Glove for Human-to-Robot Skill Transfer (RA-L 2026), SPIDER (IROS 2026), and Multisensory Continual Learning: Adapting Pretrained Visuomotor Policies to Force (2026). Previously, he was a researcher at Meta FAIR and is currently a postdoctoral researcher at Stanford University working with Shuran Song.

​Pre-Reading

​• DexterityGen: Foundation Controller for Unprecedented Dexterity (RSS 2025)

​• SPIDER: Scalable Physics-Informed DExterous Retargeting (IROS 2026)

​Topics

​• Foundation controllers for dexterous manipulation

​• Physics-informed human-to-robot retargeting

​• Learning dexterous skills from teleoperation and human videos

​• Human demonstrations for robot foundation models

​• Cross-embodiment manipulation and generalization

​Open Discussion + Q&A

​• Can human videos become the largest source of supervision for robot foundation models?

​• What is the best representation for transferring dexterous skills across embodiments?

​• How should teleoperation, video, tactile sensing, and simulation be combined for large-scale robot learning?

​• Are foundation controllers the missing ingredient for general-purpose dexterous manipulation?