CV
Education, experience, selected publications, and projects.
Contact Information
| Name | Qirui Hu |
| Professional Title | Master's Student |
| 19872217049@163.com | |
| Location | Beijing, |
| Website | https://hqr-robotic.github.io |
Professional Summary
Master’s student at the Institute of Computing Technology, Chinese Academy of Sciences, working on embodied intelligence, robot manipulation, imitation learning, and visual perception.
Education
Experience
-
Embodied Intelligence Intern
LimX Dynamics
FluxDAgger, dual-arm robot learning, DAgger pipeline, and real-robot deployment.
- Designed and implemented FluxDAgger, a model-decoupled DAgger pipeline for AgileX Piper dual-arm collaborative robots that decouples policy inference, reward modeling, data collection control, and dataset processing.
- Built synchronized multi-camera observation, online human takeover, corrective DAgger collection, and Qwen3-VL reward visualization modules in a modular ROS architecture.
- Reworked the original single-process architecture into a multi-process data collection and inference system, reducing bottlenecks during model rollout and data recording.
- Redesigned the CAN bus topology for independent four-arm control, integrated a three-layer safety mechanism, and deployed SARM reward modeling as a decoupled ROS node.
- Output: FluxDAgger project page / Code; ARM: Advantage Reward Modeling for Long-Horizon Manipulation, CVPR 2026 Workshop Poster.
2025 - present China
-
AI Algorithm Intern
Beijing Humanoid Robot Innovation Center
Teleoperation data collection, Diffusion Policy reproduction, and manipulation perception.
- Built a ROS-based teleoperation data collection pipeline with Haption and UR5e by referencing Diffusion Policy source code and Haption technical documentation.
- Reproduced Diffusion Policy, designed manipulation task scenarios, coordinated outsourced multi-task data collection, and conducted real-robot training.
- Integrated camera hand-eye calibration into the policy pipeline and explored multi-camera perspectives with BEVFormer-style representations to improve small-model generalization.
- Output: OmniD: Generalizable Robot Manipulation Policy via Image-Based BEV Representation.
2024 - 2025 Beijing, China
Awards
-
2021, 2023 National Scholarship
Ministry of Education, China
-
2021, 2023 Jinlongyu Scholarship
China Agricultural University
Comprehensive rank 1st.
-
2020, 2021, 2022, 2023 First-Prize Academic Scholarship
China Agricultural University
-
2024 Outstanding Graduate Award
China Agricultural University
Competitions
- 2022: First Prize (Beijing), National Undergraduate Mathematical Contest in Modeling, Group A.
- 2022: First Prize (Beijing), National Undergraduate Mathematics Competition, non-math category.
- 2022: Meritorious Winner (M Award, Top 7%), Mathematical Contest in Modeling / Interdisciplinary Contest in Modeling.
- 2021: First Prize (Beijing), National Undergraduate Mathematical Contest in Modeling, Group A.
- 2021: Second Prize (Beijing), National Undergraduate Mathematics Competition, non-math category.
Publications
-
2025 OmniD: Generalizable Robot Manipulation Policy via Image-Based BEV Representation
Preprint
Real-robot manipulation policy study using image-based BEV representations for improved policy generalization.
-
2026 ARM: Advantage Reward Modeling for Long-Horizon Manipulation
CVPR 2026 Workshop Poster
Reward modeling for long-horizon manipulation tasks.
-
2026 TEAR of the SUNSET: A Benchmark for Road Detection in Semi-Structured Environments
ICME 2026
Benchmark and algorithm study for road detection in semi-structured environments.
Skills
Robotics and Embodied AI: Robot Manipulation, Imitation Learning, DAgger, Vision-Language-Action Policies, Real-Robot Deployment
Perception and Learning: Visual Perception, Multi-Camera Perception, BEV Representation, Lane Detection, Object Detection
Systems: ROS, ROS2, Multi-Process Data Collection, Sensor Synchronization, Hand-Eye Calibration
Languages
Chinese : Native speaker
English : CET-6
Interests
Research: Embodied intelligence, Robot manipulation, Imitation learning, Visual perception
Projects
-
FluxDAgger
Model-decoupled DAgger pipeline for dual-arm robotic manipulation.
- Decouples policy inference, reward modeling, data collection control, and dataset processing.
- Supports synchronized multi-camera observation, online human takeover, corrective DAgger data collection, and reward visualization.
- Provides ROS nodes for sync observation, DAgger control, DAgger collection, Qwen3 reward inference, and four-arm Piper control.
-
Gesture Recognition on Horizon X3 Pi
Undergraduate thesis on efficient gesture recognition and deployment for mobile robots.
- Trained YOLOv5s on the HaGRID static gesture dataset, achieving 86.60% mAP.
- Deployed YOLOv5s on Horizon X3 Pi and improved inference from 20 FPS to 30 FPS with only 0.36% mAP loss.
- Developed ROS2 motion-control nodes integrating gesture recognition and human tracking into a mobile robot.