Harry H. Zhang

I am a graduate student in the SPARK Lab of MIT LIDS. I am extremely fortunate to be advised by Prof. Luca Carlone.

Prior to MIT, I was a MS-Research student in the CMU Robotics Institute studying Artificial Intelligence and Robotics, advised by Prof. David Held. I also worked at Amazon as an Applied Scientist II.

Prior to CMU, I earned my B.S. (2017-2021) with Honors from UC Berkeley with a major in EECS and a minor in Mechanical Engineering. During my time at Berkeley, I did research under Prof. Ken Goldberg and Dr. Jeffrey Ichnowski in AUTOLab. I maintain and curate a popular deep reinforcement learning tutorial on my Github.

Outside of school, I do quantitative finance.

Email  /  Google Scholar  /  Github

In reverse chronological order:

• Jan. 2026: H2OFlow accepted to ICLR, see you in Rio!
• May. 2025: CUPS accepted to ICML, see you in Vancouver!
• Jan. 2025: Conformalized HPE paper accepted to ICLR, see you in Singapore!
• Aug. 2023: FlowBot++ accepted to CoRL, see you in ATL!
• Sep. 2022: TAX-Pose accepted to CoRL, see you in New Zealand!
• Apr. 2022: FlowBot3D accepted to RSS, see you in NYC!
• May. 2021: Won the Warren Y. Dere Award from UC Berkeley EECS.
• Jun. 2020: Dex-Net AR got featured on VentureBeat.
• Jun. 2020: Dex-Net AR got featured on Sohu.

My current research focuses on trustworthy AI and autonomous systems. Specifically, I design algorithms for machines to learn representations for more robust real-world generalization and better certifiability. My research revolves around the theme of learning-based perception systems and robotic systems.

We introduce H2OFlow, a novel framework that comprehensively learns 3D HOI affordances — encompassing contact, orientation, and spatial occupancy— using only synthetic data generated from 3D generative models. H2OFlow employs a dense 3D-flowbased representation, learned through a dense diffusion process operating on point clouds. This learned flow enables the discovery of rich 3D affordances without the need for human annotations.

We model the noise in the process and observation model of nonlinear non-Gaussian systems as Max-Entropy Distributions (MED). We propagate the moments through the process model and recover the distribution as MED, thus avoiding symbolic integration, which is generally intractable. All the steps in MEM-KF, including the extraction of a point estimate, can be solved via convex optimization. We showcase the MEM-KF in challenging robotics tasks, such as localization with unknown data association.

We introduce CUPS, a novel method for learning sequence- to-sequence 3D human shapes and poses from RGB videos with uncertainty quantification. To improve on top of prior work, we develop a method to score multiple hypothe- ses proposed during training, effectively integrating uncer- tainty into the learning process. This process results in a deep uncertainty function that is trained end-to-end with the 3D pose estimator. Post-training, the learned deep uncer- tainty model is used as the conformity score. Since the data in human pose-shape learning is not fully exchangeable, we also pro- vide two practical bounds for the coverage gap in confor- mal prediction, developing theoretical backing for the un- certainty bound of our model.

We introduce CHAMP, a novel method for learning sequence-to-sequence, multi-hypothesis 3D human poses from 2D keypoints by leveraging a conditional distribution with a diffusion model. To predict a single output 3D pose sequence, we generate and aggregate multiple 3D pose hypotheses. For better aggregation results, we develop a method to score these hypotheses during training, effectively integrating conformal prediction into the learning process. This process results in a differentiable conformal predictor that is trained end-to-end with the 3D pose estimator. Post-training, the learned scoring model is used as the conformity score, and the 3D pose estimator is combined with a conformal predictor to select the most accurate hypotheses for downstream aggregation.

We consider the problem of estimating object pose and shape from an RGB-D image. Our first contribution is to introduce CRISP, a category-agnostic object pose and shape estimation pipeline. The pipeline implements an encoder-decoder model for shape estimation. It uses FiLM-conditioning for implicit shape reconstruction and a DPT-based network for estimating pose-normalized points for pose estimation. As a second contribution, we propose an optimization-based pose and shape corrector that can correct estimation errors caused by a domain gap.

We investigate a variation of the 3D registration problem, named multi-model 3D registration. In the multi-model registration problem, we are given two point clouds picturing a set of objects at different poses (and possibly including points belonging to the background) and we want to simultaneously reconstruct how all objects moved between the two point clouds.

We explore yet another novel method to perceive and manipulate 3D articulated objects that generalizes to enable the robot to articulate unseen classes of objects.

We conjecture that the task-specific pose relationship between relevant parts of interacting objects is a generalizable notion of a manipulation task that can transfer to new objects. We call this task-specific pose relationship "cross-pose". We propose a vision-based system that learns to estimate the cross-pose between two objects for a given manipulation task.

We explore a novel method to perceive and manipulate 3D articulated objects that generalizes to enable the robot to articulate unseen classes of objects.

We present present AVPLUG: Approach Vector PLanning for Unicontact Grasping: an algorithm for efficiently finding the approach vector using an efficient oct-tree occupancy model and Minkowski sum computation to maximize information gain.

We propose a self-supervised learning framework that enables a UR5 robot to perform these three tasks. The framework finds a 3D apex point for the robot arm, which, together with a task-specific trajectory function, defines an arcing motion that dynamically manipulates the cable to perform tasks with varying obstacle and target locations.

We present a distributed pipeline, Dex-Net AR, that allows point clouds to be uploaded to a server in our lab, cleaned, and evaluated by Dex-Net grasp planner to generate a grasp axis that is returned and displayed as an overlay on the object.

We present an algorithm to orient novel objects given a depth image of the object in its current and desired orientation.

We present an algorithm to train a robot to control free-end cables in a self-supervised fashion.

We introduce andexplore a novel method for adding safety constraints for model-based RL during training and policy learning.

10-725: Graduate Convex Optimization
16-385: Computer Vision

CS 189: Introduction to Machine Learning

EE 127: Introduction to Convex Optimization

Website template from Jon Barron