Trained a PPO posture-control policy to replace a model-based QP/IK controller; deployed on hardware and refined with real-world data, validating on three outdoor terrains unseen in training (grass, gravel, pavement).
Benchmarked three architectures (reactive, recurrent, and teacher-distilled) with identical reward and observation spaces; showed that direct RL on the recurrent policy reached deployment without two-stage distillation.
Built the platform end-to-end: mechanical and mechatronic design, a low-level base controller over CAN, and a data pipeline to onboard compute.
Research
FORLAX: First-Order Reinforcement Learning in JAX
Sep 2026 – Present
Student Researcher, Robomechanics Lab · PI: Prof. Aaron Johnson
Leading the Unitree G1 humanoid velocity-tracking and LEAP Hand in-hand reorientation tasks within a first-order RL pipeline built on differentiable simulation in JAX/MuJoCo.
Converting non-differentiable reward and termination terms into smooth surrogates and tuning first-order training.
Benchmarking against zeroth-order baselines (PPO) on sample efficiency and wall-clock time.
Self-Stabilizing Penguin-inspired Bipedal Robot for Low-Friction Terrain
Aug 2025 – Present
Student Researcher, Robomechanics Lab · PI: Prof. Aaron Johnson
Designed (SolidWorks) and built a 5-DOF penguin-inspired biped with a crank-slider leg extension and adjustable center-of-mass height; modeled its closed-chain leg mechanism in MuJoCo and NVIDIA Isaac Lab.
Swept gait parameters across four friction coefficients in MuJoCo and validated on hardware with motion capture: penguin-style torso-over-stance-leg motion with a high COM was 1.8× faster with 44% lower cost of transport than an upright human-like gait at μ = 0.12, reversing at μ = 0.45.
Extended the study to closed-loop control: trained CPG-RL and imitation RL policies in Isaac Lab that climb 6° slopes at μ = 0.1 in simulation.
→ Co-first author, "When to Waddle" · Under review, ICRA 2027 (Seoul) · arXiv:2609.21185 · Co-first-author workshop, ICRA 2026 (Vienna) · Co-first-author workshop, IROS 2026 (Pittsburgh) · Live demos, NFL Draft Tech Day 2026, Robotics & AI Discovery Day 2026, and NSF Foundational Research in Robotics (FRR) 2026
Design and Control for Mugatu: A Minimally-Actuated Steerable Bipedal Robot
Dec 2023 – Sep 2026
Student Researcher, Robomechanics Lab · PI: Prof. Aaron Johnson
Engineered a PID attitude controller for the first steerable single-motor biped, enabling directional control.
Conducted 2D contact dynamics analysis in MATLAB to investigate limit cycle stability and optimize gait efficiency.
Validated non-linear dynamics and steering manifolds using Drake and MuJoCo to bridge theoretical stability, simulation, and physical execution.
→ Live demo, Robotics & AI Discovery Day 2025
Allometric Scaling Laws for Bipedal Robots
Dec 2023 – Sep 2026
Student Researcher, Robomechanics Lab · PI: Prof. Aaron Johnson
How should a biped's design change with its size? A survey of existing bipedal robots shows mass scales as L2 rather than the isometric L3.
Mugatu (0.15 m legs) and Zippy (0.025 m legs) simulated in Drake across a 60× range of leg lengths: walking speed follows dynamic similarity (v ∝ L½), joint torque scales as mL, and feasible foot size scales linearly with leg length.
→ Under review, ICRA 2027 (Seoul) · arXiv:2603.22560 · Presented at GRC Robotics 2026, NCUR 2025, CMU MOTM 2025
Designed a rack-and-pinion gripper and parallelogram lift for a two-servo robot car as parametric CAD scripted in CadQuery: 10 printed parts at ≈83 g, 92 mm of lift travel, 28–137 mm jaw gap.
Sized a rear ballast box from a forward-tipping analysis and added plate pillars so the car could carry the pipe without rearing up or shedding its top plate.
Built a MuJoCo twin of the car and mechanism that accepts the ESP32 board's packet protocol, with a pick-and-stack demo that cancels the parallelogram arc with car motion.
Hybrid Locomotion for Unitree Go2-W in Agricultural Environments
Jan 2026 – Sep 2026
Course Project · Robotics & AI for Agriculture · Prof. George Kantor & Dr. Francisco Yandun
Implemented a PPO-based RL policy in NVIDIA Isaac Lab for adaptive wheel-leg switching on the Go2-W.
Trained robust navigation in procedurally generated off-road terrains with multi-crop row constraints.
MPPI-Based Autonomous High-Speed Drone Racing
Jan 2026 – Present
Course Project · Mobile Robots · Prof. Wennie Tabib
Developed an MPPI controller for a 17-state quadrotor model to achieve real-time, aggressive gate navigation.
Optimized trajectories in PyBullet by sampling control inputs to balance high velocity and collision avoidance.
Sim-to-Real RL: Autonomous Quadrotor Landing
Aug 2025 – Jan 2026
Course Project · Robot Learning · Prof. Guanya Shi
Synthesized a PPO controller with 4096 parallel agents; designed a curriculum-learning scheme with exp reward scaling for precision landing on moving targets (<2 cm error).
Mitigated the sim-to-real gap via domain randomization — 96% success rate on Crazyflie 2.1+ hardware.
MPC-based Gait Transition for Unitree A1
Jan 2024 – May 2024
Course Project · Optimal Control and Reinforcement Learning · Prof. Zachary Manchester
Developed a RL and LQR-based controller for seamless transitions between 3 distinct gaits within 2 seconds.
Optimized multi-contact locomotion using contact-implicit trajectory optimization in MuJoCo.
Pneumatic Artificial Muscle Suspension System
Jan 2024 – May 2024
Designed suspension system in SolidWorks and fabricated prototype with integrated pneumatic actuation.
Conducted pneumatic pressure analysis in MATLAB and ANSYS finite element analysis (FEA) to evaluate damping performance and structural integrity under dynamic loading conditions.
Publications
B. Gu*, N. Oke*, et al. "When to Waddle: A Comparative Study of Bipedal Torso-Stabilization on Low-Friction Surfaces."