Hao Jiang | 姜昊

Dexterous manipulation · Sim-to-real RL · Tactile sensing

I am a second-year PhD student in Computer Science at the University of Southern California, advised by Prof. Daniel Seita. I completed my undergraduate studies at USC, double majoring in Computer Science and Applied & Computational Mathematics.

I work on contact-rich dexterous manipulation and the sensing it requires. I combine learning with model-based structure to make robot hands work reliably in the real world, and I build the systems behind it, from tactile hands to real-to-sim twins.

I am looking for research internships for Summer 2027 in robot learning and manipulation, broadly. Feel free to reach out.
Hao Jiang

News

Selected Publications

(* indicates equal contribution, † indicates equal advising)

When Does Touch Matter?
When Does Touch Matter? Charting the Vision–Interaction Gap in Cluttered Dexterous Grasping
Hao Jiang, Luis Dominguez, Daniel Seita
In submission, 2026
project page / arXiv

A controlled study of when vision alone stops being enough for dexterous grasping in clutter. Sweeping five scene conditions of increasing occlusion and confinement charts a widening vision–interaction gap, and shows contact wrench and taxel sensing to be complementary rather than redundant.

DexMulti
Concurrent Prehensile and Nonprehensile Manipulation: A Practical Approach to Multi-Stage Dexterous Tasks
Hao Jiang, Yue Wu, Yue Wang, Gaurav S. Sukhatme, Daniel Seita
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026
project page / arXiv / code / datasets

Sample-efficient dexterous manipulation via retrieve-align-execute over object-centric skills. Achieves 66% success on concurrent multi-stage tasks with 3–4 demos per object, outperforming diffusion policy baselines by 2–3×.

SOPE
Learning to Singulate Objects in Packed Environments using a Dexterous Hand
Hao Jiang, Yuhai Wang*, Hanyang Zhou*, Daniel Seita
International Symposium of Robotics Research (ISRR), 2024
project page / arXiv / code / presentation / blog

Multi-phase RL with a displacement-based state representation for singulating target objects in clutter using a 16-DOF Allegro Hand. Demonstrates strong sim-to-real transfer and outperforms baselines in packed-environment tasks.

Other Publications

HANDFUL
HANDFUL: Sequential Grasp-Conditioned Dexterous Manipulation with Resource Awareness
Ethan Foong*, Yunshuang Li*, Hao Jiang, Gaurav S. Sukhatme, Daniel Seita
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026
project page / arXiv
UniLatent
UniLatent: Cross-Embodiment Transfer via Latent Observation Alignment
Yufei Ding*, Yi Du*, Weiduo Yuan, Hao Jiang, Dylan Goetting, Lawrence Yunliang Chen, Daniel Seita, Yue Wang
Conference on Robot Learning (CoRL), 2026
project page / arXiv / OpenReview
ENVISION
ENVISION: Embodied Visual Planning via Goal-Imagery Video Diffusion
Yuming Gu, Yizhi Wang, Yining Hong, Yipeng Gao, Hao Jiang, Angtian Wang, Bo Liu, Nathaniel S. Dennler, Zhengfei Kuang, Hao Li, Gordon Wetzstein, Chongyang Ma
arXiv, 2025
project page / arXiv

Robot Systems

Hardware and software I built and use for the research above. Videos are real robots unless marked as simulation.

Open-source LEAP Hand in-hand manipulation Yaw rotation, SO(3) reorientation and palm-down fingertip rotation of 22 objects, with PPO and a world model that reaches PPO's level ~15× faster. Includes servo model, checkpoints and deployment code. RLWorld modelsSim-to-real code ★ 45
SO(3) reorientation and deployment stack Cube reorientation to random 3D goals on the real hand (4× speed). Policies run on a GPU server and drive the hand over ROS 2, with RGB pose tracking and a live dashboard. ROS 2Pose trackingPolicy serving code
Compliant tactile hand platform xArm7 + LEAP Hand with 508 fingertip taxels, plus per-finger and wrist wrench estimates under compliant control. TactileWrench sensingCompliance
Teleoperation and data collection Glove-based teleoperation with multi-view RGB-D capture and 3D object tracking. Demonstrations from DexMulti are released as LeRobot datasets. TeleopRGB-D datasets
Real-to-sim digital twin Left: real robot, right: MuJoCo replay. Real xArm7 + LEAP Hand episodes replayed in simulation from the same commands and camera view, to measure and close the sim-to-real gap. Real-to-simMuJoCoxArm7
Actuator identification Force-gauge measurements of the LEAP Hand's servos and gearboxes, used to build a measured actuator model for training. System IDHardware

Experience

Awards

Service & Mentoring