Shuai Zhou

Shuai Zhou

shuaizho at cs dot cmu dot edu

I am an MSR 🤖 student at CMU Robotics Institute, working on dexterous manipulation with Zackory Erickson and Jeff Schneider.

Previously

I did my B.Eng. in Robotics 🤖 from SCUT (2026). Along the way I spent wonderful undergraduate years at HKUST (Guangzhou) (2026), CMU (2025), SJTU (2024), and UC Berkeley (2023), working with Haoang Li, Jiaoyang Li, Ding Zhao, and Zhongqiang Ren, where we did some fun projects spanning from classical planning to robot foundation models.

Research Collaboration:

I am interested in enabling robots with human-level physical and cognitive intelligence. I'm always excited to connect and collaborate with researchers who share a passion for robotics. Lately, I've been exploring a few fun ideas in each of the following areas and would love to find collaborators:

  • Dexterous Manipulation
  • Humanoid
  • Active Perception

Feel free to shoot me an email to hear more. I'd also love to hear your ideas and discuss what we could work on together! I'm also familiar with planning and happy to offer advice in that area.

Shuai at Bryce Canyon
Shuai at Bryce Canyon Shuai at Compton with Todd

News

  • 2026/10 HTD received Best Paper Award at IROS 2026 PWMS Workshop.
  • 2026/09 LAMP is featured by CMU RI.
  • 2026/08 Join CMU RI as an MSR.
  • 2026/06 LAMP, HTD are accepted to IROS 2026.
  • 2026/03 Start my visiting at HKUST (Guangzhou).
  • 2025/06 Start my visiting at CMU RI.
  • 2025/05 LSRP* is accepted to SoCS 2025.
  • 2024/12 LSRP is accepted to AAAI 2025.
  • 2023/08 Start my exchange at UC Berkeley.

Blogs

EgoANT: Segmentation & Annotation for Egocentric Human Video

TL;DR: Turning raw egocentric human video into time-aligned subtask labels for robot learning, using only open-weight, self-hostable VLMs. EgoANT finds action boundaries with a whole-episode coarse cut over timestamped contact sheets followed by local refinement, then picks the best of several candidate descriptions for each segment. On WGO-Bench (HomER), it more than doubles segmentation F1 over wrist-speed heuristics (0.095 → 0.203) and approaches a Gemini-based pipeline (0.227), at roughly $1–3 per video-hour.

report code

Publications

( / )

Topics: / / /

* denotes equal contribution.

UniWAM: Unified World-Action Model

Under Review

TL;DR: VLAs inherit strong reasoning, but action-only supervision provides limited grounding in world dynamics; world-action models capture dynamics but remain limited in semantic reasoning and generalization. UniWAM unifies a VLM reasoner, a video world model, and an action expert in one model, co-trained on ~10k hours of human and robot data annotated with our EgoANT pipeline. It sets SOTA on LIBERO (99.2%) and LIBERO-Plus (92.6%), and human data keeps paying off as scale grows.

paper project code

ActiveScale: Scaling Active Perception for Robots across Model, Data, and Hardware

Under Review
ScalingH2R Workshop, CoRL 2026

TL;DR: A full-stack recipe for scaling active perception in robot foundation models, spanning model, data, and hardware. ActiveScale pairs a pose-aware VLA with the first 1,000-hour human–robot active-perception training mix (over 5× prior efforts) and AMP, a single-operator mobile platform with an actuated head camera. Robots learn to look for hidden or out-of-view objects, raising real-world success from 30% to 70% over π0.5.

paper project code

From Discrete Plans to Real-World Execution: A World-Model-Driven Framework for Execution-Aware Multi-Agent Path Finding

Under Review

TL;DR: MAPF plans that are optimal under planning assumptions are often far from optimal in real-world execution. We close this planning–execution gap with the first learned world model of MAPF execution, which predicts how a plan will actually unfold and steers both planning and scheduling. It cuts deadline misses by up to 27% in simulation with up to 300 agents and real-robot execution time by 15.3%.

paper

EgoX: Egocentric Cross-Embodiment Manipulation with Embodiment Dreaming

Under Review
ScalingH2R Workshop, CoRL 2026
Dex Workshop, RSS 2026

TL;DR: One manipulation policy that learns from humans and very different robots alike. EgoX unifies human, humanoid, wheeled, quadruped, arm, and tool-use demos into one egocentric format and pretrains a modular Transformer that “dreams” each body’s future state. Cross-embodiment pretraining raises task progress from 0.63 to 0.80, with the biggest gains out of distribution.

paper project

LAMP: Long-Horizon Adaptive Manipulation Planning for Multi-Robot Collaboration in Cluttered Space

IROS 2026

TL;DR: In clutter, an object path that looks feasible may leave robots no room to make contact and manipulate it. LAMP plans the object, the robot team, and their contacts together, verifying with a learned manipulation model during search whether each step is actually executable, and replans in seconds when execution drifts. It solves 88–100% of densely cluttered tasks where prior methods mostly fail, including assembling a 9-object “IROS” sequence end to end.

paper project code news

HTD: Learning Versatile Humanoid Manipulation with Touch Dreaming

IROS 2026
PWMS Workshop, IROS 2026, Best Paper Award
WCBM Workshop & SemRob Workshop, RSS 2026

TL;DR: Humanoids that manipulate by feel, not just sight. HTD treats touch as a core modality and learns JEPA-style, predicting future hand forces and tactile latents alongside actions, on top of a robust whole-body controller and VR teleoperation. It nearly doubles success on five contact-rich real-world tasks (+90.9% relative), from towel folding to tea serving.

paper project code

Loosely Synchronized Rule-Based Planning for Multi-Agent Path Finding with Asynchronous Actions

AAAI 2025

TL;DR: Real robots don’t move in lockstep, yet most scalable MAPF solvers assume they do. LSRP is the first to scale MAPF with asynchronous actions to hundreds and up to a thousand agents, by loosely synchronizing a fast rule-based planner. Under the same runtime limit it handles an order of magnitude more agents than prior methods, at about 25% longer makespan.

paper project code

Experience

Miscellanea

  • I play badminton and was an athlete. Now I am also a rock-climbing rookie.
  • I am from Shenzhen, China. I've also lived in Guangzhou, Shanghai, Berkeley, Los Angeles, Pittsburgh, Philadelphia, Taipei, and Hong Kong. I've done some solo backpacking and prefer staying for a while to get to know the locals.
  • My life is all about soul, jazz, and cinema. I'd love to hear your recommendations.