Cunxi (Jimmy) Dai

I'm a Ph.D. student at CMU RI advised by Prof. Guanya Shi. Previously, I was a MSR student advised by Prof. Ralph Hollis, Previous to CMU, I worked with Prof. Zhenzhong Jia, Prof. Mingming Zhang at SUSTech and Prof. Hugh Herr at MIT Media Lab.

I care about robots 🦾 that physically interact with the world and assist humans 👫 in all sorts of ways. My research aims to make contact scalable — developing algorithms and data pipelines that let robots master contact-rich interaction, from humanoid loco-manipulation to dexterous manipulation, and adapt intelligently to real-world uncertainty.

Email  /  CV  /  Google Scholar

TL;DR: CARD is a simulator- and object-free framework that retargets human demonstrations to robot actions by explicitly preserving contact force targets, unifying video, motion capture, and tactile-glove data through a shared hand-contact representation.

TL;DR: Ariadne learns the underlying state-space manifold from state-only observations and uses its geometry to construct feasible trajectories, generalizing to unseen start-goal pairs without any expert trajectory data.

TL;DR: ENPIRE lets coding agents construct real-world robot environments and autonomously iterate on policy improvement, enabling robot policies to self-improve on dexterous manipulation tasks without human algorithm engineering.

TL;DR: PGDG is a data generation framework with zero-shot curation that expands a single demonstration into a compact dataset of physically plausible recovery behaviors, enabling robust contact-rich bimanual policy learning.

TL;DR: BFM-Zero is a behavioral foundation model that learns a shared latent space over motions, goals, and rewards via unsupervised RL, letting a single humanoid robot policy be prompted to perform diverse whole-body tasks in the real world.

TL;DR: A holistic framework for dynamic interactive navigation that leverages learned representations for rapid adaptation to object dynamics, and utilizes model-based prediction to inform decision-making.

TL;DR: A bi-level MPC framework that enable non-preodic upper-body contact to achieve more agile locomotion.

TL;DR: Wheelchair maneuvering with Ballbot that balance on a single-spherical wheel.

TL;DR: Minimalist actuator realization of a wheel-leg transformable robot with tendon-driven transformation wheel module.

TL;DR: Robust design of a wheel-leg transformable robot for high-payload context.

TL;DR: Knee joint misalignment reduction with an RPR mechanism-based device.

Selected Projects

Template from Jon Barron.