Hi, I’m
Zheyu Zhuang · Postdoctoral researcher at KTH
Robots that adapt.
And understand the task.
I study how robots learn efficiently, adapt to new situations, and follow human intent through language.
My work connects perception, imitation learning and real robots. I’m interested in how they understand long, multi-step tasks: what matters now, what comes next, and when to change course.
I work with Professor Danica Kragic at KTH. My PhD with Professor Robert Mahony at ANU focused on learning-based visual control.
A selection of projects, from learning where to look to taking a skill into a new setting.
Attention for manipulation
Seeker CoRL 2026
Seeker learns where to look through action prediction alone, without spatial labels. The same attention guides zoom, background augmentation and 3D point filtering.
60% versus 20% for the strongest baseline, averaged across lighting and background changes.
Attention-guided training images
New backgrounds at evaluation
Task-aware efficiency
FocusPool CoRL 2026
Six recorded moments, looping through the task’s changing focus.
FocusPool selects shallow visual features according to the task’s progress. Knowing where to look makes a small encoder enough for effective control.
Using only the shallow L1 features.
The L2 model (5.8%) reaches comparable RVT-2 performance with roughly half the demonstrations: 50 versus 100 in the simulation comparison.
Skill transfer & spatial generalisation
PALM RA-L 2026
Settings can look different overall, yet look similar near the gripper. PALM pairs this local visual similarity with proprioceptive representation alignment to reuse manipulation skills across workspace positions, camera viewpoints and robot embodiments.
Needed to transfer across positions, viewpoints and robot embodiments.
System integration
I integrate hardware and software through ROS, connecting perception, sensor synchronisation, calibration and real-time control in complete robot systems.
Human–robot collaboration
I co-developed GoferBot, a vision-based human–robot interaction system for collaborative furniture assembly. It combines human action recognition, visual servoing and handover behaviour so the robot can coordinate naturally with a person.
Warehouse picking · ACRV team
I contributed as a member of the ACRV team that built Cartman, combining custom hardware with robotic perception for warehouse picking. The system won the 2017 Amazon Robotics Challenge, and the team leader later spun this work out into Lyro Robotics.
I work with PhD researchers on questions, experiments and clear communication, with publications at CoRL, RA-L and ICRA.
Six courses tutored at ANU, plus hands-on Robotic Vision Summer School workshops I designed and led. The workshop material later became part of teaching at Monash University.
Associate Editor for IROS 2025 and 2026 and ICRA 2027. Reviewer for IEEE RA-L, IEEE T-RO, RSS, CoRL, ICLR, ICRA and IROS.
Usually outside.
Often off-grid.
A clear head, a loud survival instinct, and plenty of pauses to rethink my approach.
Days without a shower. A clear head, big views and being off-grid make up for it.
Plenty of trips without a catch. One fish is enough to keep me going.