# Junhua (Harry) Huang > Robot learning and robotics systems. M.S. in Electrical and Computer Engineering at UCLA (expected 2027); Student Researcher at the Vision and Autonomy Intelligence Lab (VAIL), UCLA, with Prof. Bolei Zhou. Los Angeles, CA. Open to full-time roles and PhD positions in robot learning and robotics, starting 2027. Junhua (Harry) Huang works where robot learning meets real hardware: cross-embodiment data, human-in-the-loop training, and on-robot deployment. He is first author of ChairNav, a data-efficient cross-embodiment pipeline for outdoor, map-free wheelchair navigation on a Permobil M3, and co-author of Cue the Flow (CoRL 2026), ToolSmith (Findings of EMNLP 2026) and S3 (ICML 2026 Spotlight). His experience covers the whole robot stack, from reverse-engineering a wheelchair's R-net CAN interface, nonlinear system identification, 3D-printed prototypes and sensor mounts, and semi-automated RGB-LiDAR calibration, to quantized TensorRT model inference on a Jetson AGX Orin, edge-host VLA deployment on a quadruped mobile manipulator, eye-gaze interaction with Meta Aria glasses, vLLM model serving and city-scale Gaussian Splatting. He prototypes quickly, deploys on real robots, and improves systems from real-world use, for example human-in-the-loop post-training from rider takeovers. ## Links - [Homepage](https://harry-junhua-huang.github.io/): about, work, papers, education - [Drive the wheelchair model](https://harry-junhua-huang.github.io/play/wheelchair/): interactive model of the ChairNav power wheelchair, identified from its calibration drives - [CV](https://harry-junhua-huang.github.io/cv/): full text CV - [Résumé (PDF, 1 page)](https://harry-junhua-huang.github.io/files/Junhua_Huang_Robotics_Resume.pdf) - [Publications](https://harry-junhua-huang.github.io/publications/) - [LinkedIn](https://www.linkedin.com/in/junuha-huang/) - [GitHub](https://github.com/Harry-Junhua-Huang) - Email: junhuahuang1218@gmail.com ## Experience - **Student Researcher, Vision and Autonomy Intelligence Lab (VAIL), UCLA**, Prof. Bolei Zhou (Jan 2026 – Present) - ChairNav (first author): cross-embodiment training-data pipeline that adapts 170 h of source delivery-robot teleoperation video to a Permobil M3 power wheelchair with synthetic trajectories; imitation learning on 3.19 h of human driving demonstrations plus human-in-the-loop post-training. The post-training policy drove 84.9% of driving time autonomously over 7.8 km in four outdoor drives (116.2 of 136.9 min; a share of driving time, not a task success rate). - Wheelchair systems: reverse-engineered the R-net CAN interface with feedforward and residual compensation; identified nonlinear wheelchair dynamics for physics-based trajectory simulation; sped up ChairNav model inference on a Jetson AGX Orin from 0.6 to 20 Hz with model quantization and TensorRT (model inference rate, not the full control loop; the paper's deployment runs the model at 10 Hz). - Broader VAIL engineering: Isaac Sim and UrbanSim; semi-automated RGB-LiDAR calibration; conditional auto-exposure; 3D-printed mounts; automated robot-data reconstruction, visualization and filtering; GPS/IMU/odometry localization; multi-IMU and high-precision GPS data collection. - Cue the Flow (CoRL 2026): deployed VLM grounding and a frozen π0.5 VLA on an AgileX PiPER arm, with a Jetson handling camera processing and compression and bridging to remote inference in a synchronous control loop; integrated the arm, power, communications and streaming on a DEEP Robotics M20 Pro; built semi-automated demonstration collection and cleaning. - Meta Aria Gen 1 x M20 Pro: shared perception for online mapping and target navigation. - Meta Aria Gen 2: streaming, eye gaze, head pose, image depth and hand gestures to infer target direction, distance and user intent. One-day gaze-to-goal study (Sep 2026; one wearer; on-device recordings analysed offline): after a 5-point calibration, gaze error p50 1.6-2.0 degrees and floor-goal error 0.1-0.25 m at 1.5-4 m with the chair parked, checked against AprilTag ground truth. In this study, distance came from the calibrated gaze ray intersected with the floor plane (the glasses' own depth/vergence output proved unusable); not yet tested live-streamed or while moving. - Also: Booster T1 humanoid data collection; Unitree Go2 navigation host software. - **Student Researcher, UCLA NLP Group**, Prof. Kai-Wei Chang (Dec 2025 – Mar 2026) - Co-developed ToolSmith: 264 verified tools generated from a 9,174-query pool through tool synthesis and verification; Qwen3-VL-8B MMSearch accuracy 51.46% -> 55.56%. - **Student Researcher, Seizure Lab, ECE Department, UCLA** (Aug 2025 – Oct 2025) - Built a clinical seizure-semiology dataset (10,000 records, 8 features) contributing to S3; deployed 3 VLMs with vLLM asynchronous inference, 0.3 -> 5 requests/s (16.7x). - **Student Researcher / Research Assistant, University of Rochester**, Prof. Chenliang Xu's lab and Prof. Jiebo Luo's lab (2024 – 2025) - Prof. Chenliang Xu's lab: first author of Semantic Visually-Guided Acoustic Highlighting with Large Vision-Language Models (VLM-guided audio remixing; arXiv 2026); co-speech gesture generation, contributing to GestureLSM (ICCV 2025). - Prof. Jiebo Luo's lab: diffusion-based image generation (Stable Diffusion image-to-image from skeletons; preference optimization with DPO and LoRA); contributed to the Video-LMM Post-Training survey. - **Machine Learning Intern, Farsee2 Technology** (Hybrid; Jul 2024 – Sep 2024) - About 40% fewer floating artifacts in city-scale Gaussian Splatting with incidence-angle weighting and depth/normal supervision; multi-machine depth-generation and preprocessing scripts, integrated into the product pipeline. ## Education - University of California, Los Angeles: M.S. in Electrical and Computer Engineering, expected 2027 - University of Rochester: B.S. in Computer Science and B.S. in Business, May 2025; GPA 3.82/4.00; Highest Distinction, Cum Laude; Dean's List (8 semesters); Schwartz Discover Grant; TA for CSC171/172 ## Selected work - [ChairNav: Cross-Embodiment Pre-training and Personalization for Long-Horizon Wheelchair Navigation](https://arxiv.org/abs/2610.24012). J. Huang*, Z. Liu*, et al. Under submission, 2026 (*equal contribution). [Project page](https://vail.cs.ucla.edu/ChairNav/) - [Cue the Flow: Steering Flow-Matching Policies for Open-World Delivery Manipulation](https://arxiv.org/abs/2609.38989). H. Wang et al. (incl. J. Huang). CoRL 2026. [Project page](https://hatchetproject.github.io/delivery_steer/) - [ToolSmith: An Automatic Tool-Generation Framework for Multimodal Search Agents](https://threesr.github.io/toolsmith/). R. Sun et al. (incl. J. Huang). Findings of EMNLP 2026. - [Seizure-Semiology-Suite (S3): A Clinically Multimodal Dataset, Benchmark, and Models for Seizure Semiology Understanding](https://arxiv.org/abs/2605.21852). L. Zhang et al. (incl. J. Huang). ICML 2026, Spotlight. - [GestureLSM: Latent Shortcut based Co-Speech Gesture Generation with Spatial-Temporal Modeling](https://arxiv.org/abs/2501.18898). P. Liu et al. (incl. J. Huang). ICCV 2025. - [Semantic Visually-Guided Acoustic Highlighting with Large Vision-Language Models](https://arxiv.org/abs/2601.08871). J. Huang, C. Huang, C. Xu. arXiv 2026. - [GestureFAR: Streaming Co-Speech Gesture Generation with Flow Autoregression](https://arxiv.org/abs/2609.21576). P. Liu et al. (incl. J. Huang). arXiv 2026. ## Skills & platforms - Robot learning: imitation learning, preference learning, reinforcement learning, flow matching, VLA, VLM, cross-embodiment transfer, human-in-the-loop learning - Robotics systems: system identification, physics-based simulation, Isaac Sim, UrbanSim, ROS, ROS 2, CAN bus (R-net), feedforward and residual compensation, TensorRT, model quantization, edge deployment, SLAM, RGB-LiDAR calibration, GPS/IMU/odometry localization, 3D printing, 3D reconstruction, Gaussian Splatting - ML systems: vLLM serving, asynchronous inference, multi-machine data pipelines, multimodal agents - Programming: Python, PyTorch, Linux - Robots: Permobil M3, AgileX PiPER, DEEP Robotics M20 Pro, Booster T1, Unitree Go2 - Wearables: Meta Aria Gen 1, Meta Aria Gen 2 - Edge computing: NVIDIA Jetson Xavier NX, NVIDIA Jetson AGX Orin ## Contact Recruiters and sourcing agents are welcome to reach out at junhuahuang1218@gmail.com.