PROJECTS / INTERACTIVE SYSTEMS

Research ideas tested through working systems.

Selected implementations across robotic manipulation, autonomous navigation, legged reinforcement learning, and control. The Go2 work is an ongoing independent exploration rather than a completed research result.

View CV
01 / 2026

Gripper-aware VLA / MiGA

Studied how end-effector morphology changes feasible manipulation strategies and designed a gripper-aware VLA framework with multi-gripper tokens, adapter routing, and a MiGA-style dataset organization for cross-gripper policy learning.

VLApi0.5Gripper EmbodimentAdapter Tuning
02 / 2026

GCA-Bench

Designed a benchmark framing for complex grasping tasks that require semantic reasoning, scene understanding, multi-stage trajectory execution, and task success evaluation across simulated and real-world settings.

BenchmarkComplex GraspingVLA EvaluationSim-to-Real
03 / 2026

Legged Locomotion with PPO in MuJoCo

Ongoing Independent Project

Exploring PPO-based legged locomotion in MuJoCo, with a focus on reward design, curriculum learning, proprioceptive observations, and policy evaluation.

PPOMuJoCoLegged LocomotionCurriculum Learning
04 / 2025

LeRobot SO-101 Implementation

Groundwork for an autonomous magnetic capsule endoscopy arm at Jifu Medical. Led the teleoperation and data collection pipeline on SO-101 dual-arm robots with LeRobot (ACT / ALOHA): synchronized multi-view recording and joint-state / action logging, with 50+ episodes recorded and validated. This was the first step toward a VLA-style system that reads patient posture and capsule images to steer the capsule without a doctor driving the arm.

Imitation LearningLeRobotRobot ManipulationData Collection
Solve its IK live ↓Inspect the arm →Pipeline figure →Capsule system →
05 / 2025

Self-Balancing Wheeled-Legged Robot

Independent Build

Built end-to-end on the open-source StackForce bipedal wheeled platform: an ESP32-S3 with an MPU6050 IMU, two BLDC hub motors with magnetic encoders, and four servos driving five-bar legs. Firmware in embedded C++: five-bar inverse kinematics, a cascaded PID balance loop (speed → pitch → wheel torque), yaw-rate control, and roll-based leg-height adaptation for rough terrain, under SBUS remote control.

C++ESP32-S3Cascaded PIDFive-bar IKMPU6050 IMUBLDC + servos

LIVE SYSTEM / 01

SO-101 arm · inverse kinematics

The SO-101's shoulder-lift, elbow and wrist axes are parallel, so in its vertical plane the arm is a three-link chain. Click to set a target or drop obstacles; a particle-swarm solver finds joint angles that reach it without collisions.

Loading robotic arm simulator…

LIVE SYSTEM / 02

Inverted-pendulum control

A cart-pole with one or two links: energy-shaping swing-up, a pole-placement balance controller, and supervisory switching between them. Drag the cart or press Push to disturb it and watch it recover.

View source on GitHub ↗
Loading control system…

LIVE SYSTEM / 04

Bipedal wheeled robot — balance playground

A simulation of my two-wheeled, two-legged robot (ESP32-S3, MPU6050, two BLDC hub motors, four servos on five-bar legs), driven across an endless skatepark of rolling waves, side banks, humps and a pump track. It leans forward and back to balance as it speeds up, brakes and climbs, and its legs change height to stay level on banks and lean into turns.

Take it apart

Loading exploded view…

Drive it

Loading wheel-legged robot playground…