Peiyang (Ben) Li
Student
Education
2 entriesBachelor of Science in Computer Science & Mechanical Engineering
Northwestern University ·
- GPA: 4.0 / 4.0
- Freshman Dean’s List — High Honors
IBDP Candidate
Shanghai Pinghe School
Graduated 2025
- Graduated on the Principal’s List (top 5% of class)
Experience
3 entriesUndergraduate Research Assistant
Oct 2025— Now
- Built a DW1000-based localization system with three-anchor triangulation, achieving cm-level positioning at 30 Hz update rate. Solved concurrent SPI bus contention with IMU through interrupt-driven mutex, providing a low-cost alternative to expensive motion capture systems.
- Co-developed modular robot hardware and RL control with PhD researchers. Designed novel wheel-leg modules and trained locomotion policies (PPO) across diverse morphologies in MuJoCo. Developing wheel-leg hybrid modules for multi-modal locomotion.
- Received $4,000 in Northwestern research funding to continue this work through the summer; selected by the Hung-Farinelli family as one of the fund’s featured projects.
Fig. 01Reconfigurable legged machines in the wildYouTube Robotics Engineer
Sep 2025— Jan 2026
- Built fully local inference pipeline on Nvidia Jetson with TensorFlow, achieving <20ms per-frame latency with zero cloud dependency. Delivered as a paid monthly subscription service to warehouse clients.
- Trained recognition system (YOLO detection, MobileViT super-resolution, OCR, QR/barcode) achieving 98%+ item accuracy on 2k-image custom dataset. Integrated LiDAR for 3D spatial awareness in cluttered environments.
- Developed multi-edge device orchestration with AWS IoT Core for distributed pick-and-place. Integrated LLM-based task planner for intelligent error prevention and autonomous operation.
Fig. 02Warehouse deployment, on-device camera feed Research Intern (Algorithm)
Differential Robotics
May 2025— Aug 2025
- Designed sparse-to-dense reward curriculum for obstacle avoidance, improving training convergence 3× over baseline PPO in cluttered environments in IsaacLab.
- Deployed real-time end-to-end vision model for high-speed drone maneuvers through near-vertical frames at 4+ m/s. Achieved 95% success rate with MobileNetV3 optimized via TensorRT (<5ms latency). Successfully transferred to physical hardware.
Projects
10 entriesWheel-Legged Modular Robot
Jul 2026— Now
- Building an 8-module wheel-legged robot for the lab’s modular robot platform: 3-DoF legs on driven wheels, each module an MG4010 motor with its own ESP32 and IMU.
- Built the RL training pipeline from MuJoCo/mjlab to Isaac Lab (RSL-RL PPO, GRU policies) at 2.26M physics steps/s across 32k parallel environments; policies cross slopes, steps and rough terrain in simulation.
- Brought up the hardware: calibrated six leg modules to within 3 mm and wrote a 50 Hz control daemon with watchdog and E-stop for sim-to-real deployment.
- Traced sim-to-real gaps to the plant model (measured wheel mass, a unit bug capping wheel speed near 25%) and revalidated every policy against the corrected model.
Fig. 03Simulation: 12 locomotion tasks on the full CAD model Evolved Robot Manipulators, Sim-to-Real
2026
- Built the real-world and sim-to-real side: designs evolved in simulation were printed, assembled and deployed zero-shot.
- On hardware: picked 24 unseen objects (8 of 10 attempts on average), hung from a bar under a 500 g load, and pulled a cube out of a pipe.
TestFig. 04Real-world tests of evolved designs, deployed zero-shotEvolved two-finger rotator turning a Rubik's cube End-to-End High-Speed Drone Navigation System
2025
DemoIROS 2025Science RoboticsICRA
- Led the full sim-to-real deployment of the paper’s end-to-end sensorimotor policy, enabling a physical drone to traverse 85° tilted frames at 4–6 m/s with 95% success and high repeatability; the resulting demo was exhibited at IROS 2025.
- Optimized MobileNetV3 + DeepLabV3+ on Jetson Xavier NX via TensorRT, achieving ~3 ms inference and >60 Hz closed-loop control.
- Deployed models developed in Precise Aggressive Aerial Maneuvers with Sensorimotor Policies (Science Robotics) and Whole-Body Control Through Narrow Gaps from Pixels to Action (ICRA).
Frame tiltTakeoffFig. 05Flight tests: pick a frame tilt and a takeoff pointRun 01 · 80° · 8 recorded Dexterous Hand Object Reorientation
- Assembled and calibrated a 16-DoF dexterous hand platform; built a PyBullet digital twin for joint-level control tests, policy debugging, and hardware-safe command validation.
- Trained Isaac Lab object-reorientation policies with randomized object poses, dynamics, and actuator parameters, then deployed them on the physical hand and iteratively tuned sim-to-real reliability.
Fig. 06Reorienting a cube on the physical hand ME 433 Haptic Feedback Joystick
- Designed and built a two-axis force-feedback joystick as the final class project for Northwestern’s ME 433: Advanced Mechatronics.
- Integrated modified servos, magnetic encoders, force sensing, and custom STM32/Pico 2 W electronics.
- Implemented a 500 Hz closed-loop controller that renders programmable textures and forces through the joystick.
Fig. 07Two-axis force-feedback joystick F1TENTH Autonomous Driving
- Built and deployed a ROS 2 autonomy stack on a physical F1TENTH car, integrating LiDAR SLAM, localization, and RViz-based debugging.
- Combined A* global planning with MPPI local trajectory tracking and Ackermann drive control for closed-loop navigation.
Zombies VS Plants
2026
- Built a real-time PvP strategy game inspired by Plants vs. Zombies, adding playable plant/zombie sides, local 1v1 gameplay, animated sprites, sound effects, and AI opponents trained from recorded gameplay logs.
- Integrated a computer-vision physical board interface that detects real-world game pieces and maps their type/location onto the in-game grid; built in 24 hours at WildHacks 2026.
Fig. 08Demo video, WildHacks 2026YouTube Beatbox-to-Drum Generation
2026
- Developed a beatbox-to-drum generation system that converts vocal percussion recordings into symbolic 9-class drum patterns using a flow-matching-based generative model.
- Built an audio preprocessing/model training pipeline, MIDI/audio tooling, and a human-feedback review loop for ranking outputs, training reward models, and iteratively improving rhythmic alignment and musical realism.
End-to-End RL Drone Navigation System
2025
- Built an end-to-end navigation pipeline in IsaacLab using deep reinforcement learning (PPO), mapping high-dimensional sensor inputs (depth images + proprioceptive states) directly to low-level motor actions.
- Implemented curriculum learning & domain randomization (obstacle density, dynamic disturbances, sensor noise) to improve generalization and enable sim-to-real transfer.
- Designed hybrid reward shaping: integrated Dijkstra-based path priors, directional distance rewards, ESDF collision penalties, and smoothness constraints to stabilize long-horizon flight.
Low-Cost UWB Localization for Rollbot
2025— 2026
- Built DW1000 localization with concurrent SPI bus sharing (UWB + IMU) via interrupt-driven hardware mutex for stable multi-device communication.
- Rewrote TWR ranging protocol, increasing update rate to 30 Hz and improving accuracy via Kalman-filtered multi-anchor triangulation.
- Delivered cm-level autonomous navigation for Rollbot (single-actuator spherical robot), offering a $50 embedded alternative to $10k+ OptiTrack systems.