About
I'm Junkai Huang, an undergraduate at Southwest Jiaotong University (211) studying Electronic Information Engineering — GPA 3.78/4.0 (92.02, ranked 12/80). President of the university robotics association, and embedded lead for the RoboMaster competition team.
My work sits at the intersection of real-time embedded systems and embodied AI. On the hardware side, I build firmware for STM32 microcontrollers, design PCBs, and develop motor control systems. On the AI side, I train and deploy vision-language-action models on real robot hardware — from data collection to sim-to-real transfer to edge deployment on RK3588.
I've published an EI conference paper as first author and contributed to a MICCAI 2025 CREATE Workshop paper as fourth author. I'm also second inventor on a patent application accepted for processing, and have won 9 national-level competition awards including a RoboMaster National 1st Prize.
ISTJ-T — systematic, detail-oriented, and allergic to half-finished work.
What I'm Working On
- ▸Tactile-augmented diffusion policies for contact-rich robotic assembly
- ▸Bilateral 3×3 FSR tactile arrays for contact and slip detection
- ▸Continued development of the RoboMaster robot platform
- ▸Sim-to-real transfer and TeleOp data collection pipelines
Education
Key courses: Engineering Mathematics (94), Circuit Analysis & Design (96), Power Electronics (97), Communication Principles (98), Embedded Systems (97).
Experience & Research
- ▸Independently integrated a mobile base, 7-DoF arm, gripper, RGB fisheye and depth cameras, Jetson control node, and inference server into a robot learning system.
- ▸Designed a lerobot-Uarm arm with matching kinematics for leader–follower TeleOp data collection; completed Diffusion Policy training, inference, and real-robot deployment.
- ▸Achieved approximately 50% success with a vision-only policy on picking up a nut and placing it on top of a screw.
- ▸Built bilateral 3×3 FSR tactile arrays to study contact and slip detection and action correction for fine assembly. Experiments are ongoing.
- ▸Built an embodied AI system combining natural language understanding and robot control using Dobot CR5 arm + AG-95 gripper + dual D435 depth cameras.
- ▸Collected 100+ demonstrations, built LeRobot-format dataset, LoRA fine-tuned π0 VLA model.
- ▸Voice-to-text + camera input → ΔJoint output (6-axis + gripper). Deployed via TCP after Isaac Sim validation.
- ▸"Transformer Pruning and Optimization for Embodied AI: An Embedded Implementation on RK3588"
- ▸Proposed HALSP structured compression: assess sublayer sensitivity and redundancy, prune rows, columns, and blocks for hardware efficiency, then compensate the remaining weights.
- ▸Evaluated the speed–accuracy tradeoff on RK3588; achieved <100ms inference latency with >80% task accuracy at the operating point selected for real-time use.
- ▸"A method, system and device for video anomaly object insertion and automatic mask generation"
- ▸Independently designed a pipeline that turns sparse keyframe annotations into continuous anomaly videos and per-frame pixel masks, using trajectory smoothing and edge blending. Improved annotation efficiency by approximately 20×.
- ▸Infantry (swerve drive): APP/Module/BSP layered C++ firmware on STM32H7. Self-developed MA600 hollow-shaft encoder with field compensation (<0.1° error). Gyroscope-closed-loop gimbal with MIT control — <100ms step response to 0.1 rad, near-zero steady-state error.
- ▸Engineer (7-DoF arm + mecanum): ROS2 + MoveIt2 + IKFast (4ms/solve) + KDL on upper computer. USB CDC with DMA ping-pong buffers + FreeRTOS CRC task. Self-built gravity-compensated teach pendant using absolute encoders.
- ▸Results: National 1st Prize (3rd nationally, 1st Southwest) · Alliance Match 3rd Place · Engineer National 2nd Prize — 6 national awards total.
- ▸"Anti-Forgetting Test-time Adaptation for Robust Medical Image Analysis under Distribution Shift"
- ▸Prepared, cleaned, and split EyePACS, HyperKvasir, and CVC datasets. Built controlled out-of-distribution test data with 11 corruptions across 4 categories and 5 severity levels, preserving image–label and image–mask alignment for reproducible TTA comparisons and ablations.
Awards
Tech Stack
Embedded
Robotics
AI / ML
Hardware
Software
Why Embedded Systems & Embodied AI
I like building things that exist in the physical world. Writing code that makes a motor spin precisely, a robot arm reach for an object, or a gimbal hold steady while the chassis underneath slides sideways at full speed — that feedback loop between software and physics is what keeps me engaged. Embedded systems are where every microsecond matters and every byte counts.
Embodied AI is the natural extension: giving machines the ability to perceive, reason, and act in unstructured environments using learned policies rather than hand-coded rules. I want to work at the intersection of real-time control, hardware design, and machine learning — and I'm building toward that one project at a time.