Ge Zhang张哿

Robot learning, and the data behind it.

I'm a founding member of an early-stage startup in stealth, working on robot learning.

Previously, I was an Applied Scientist at the Robotics and AI Institute (formerly Boston Dynamics AI Institute), where I trained VLA models and diffusion policies on data from UMI-style devices, and built the data pipeline and robot control stack around them. Before that, I shipped the first end-to-end visuomotor picking policy to production at RightHand Robotics.

I build learning-based systems that let robots manipulate the physical world. I focus on (a) visuomotor policy learning and (b) the data collection, curation, and evaluation pipelines behind it.

Photo of Ge Zhang

Ongoing · started at Robotics and AI Institute, 2025

Data Quality Evaluation for Human Demonstrations

A curation pipeline that decides which handheld demonstrations are worth training on, using motion heuristics, chunk-level kinematic replay in simulation, and policy-based consistency scoring. It improved success by 20% on high-accuracy tasks, and the most consistent half of a multi-collector dataset matched the full set.

Write-up

Robotics and AI Institute, 2025 – 2026

Robot-Free Data Collection: UMI Pipeline and Residual SLAM

Turning 200+ hours of handheld UMI demonstrations into robot policies. I trained a residual model that corrects pose-tracking error (30% lower on seen tasks, 15% on unseen), lifting delicate-task success from 20% to 45%. Also: why fast closed-loop policies can be more precise than their own data.

Write-up

RightHand Robotics, 2023 – 2024

Learned Visuomotor Picking in Production

Replaced a segmentation-based suction heuristic with an end-to-end policy trained on production logs, evaluated with a tier-based benchmark I designed. Deployed with 8% higher pick success on unseen items and 30% fewer trash picks.

Write-up

arXiv preprint, 2021

StrobeNet: Category-Level Multiview Reconstruction of Articulated Objects

Ge Zhang, Or Litany, Srinath Sridhar, Leonidas Guibas

WebpagePaperWrite-up

arXiv preprint, 2019

Towards 3D Human Shape Recovery Under Clothing

Xin Chen, Anqi Pang, Yu Zhu, Yuwei Li, Xi Luo, Ge Zhang, Peihao Wang, Yingliang Zhang, Shiying Li, Jingyi Yu

PaperWrite-up

ShanghaiTech University, 2019

Guidance Waist Pack for the Visually Impaired

A wearable aid that fuses RGB-D, GPS, IMU, and audio on an embedded board for navigation and object detection.

VideoWrite-up