ID// R0B0T4RL ROLE// Robotics & RL Engineer BASE// Maryland, USA YEAR// 2026

Howard H.
Cho.

I teach machines how to move. Reinforcement learning for legged robots, manipulators, and aerial systems — bridging the gap between Isaac Lab simulations and real hardware in the field.

View Work → GitHub ↗ LinkedIn ↗
01 / About
Howard H. Cho
FILE: profile.jpg STATUS: ACTIVE

A robotics and reinforcement learning engineer working at the intersection of simulation, control, and real hardware.

My focus is closing the sim-to-real gap. I design RL environments in Isaac Lab, train policies with PPO/GRPO, and deploy them onto quadrupeds, manipulators, and aerial platforms — the kind of work where the training curve is half the story and the other half lives on a robot.

The foundations come from graduate study at the University of Washington, where I worked through linear and nonlinear control, convex optimization, machine learning, and computer vision before pulling them all into the same pipeline.

Hands-on with the full stack: NVIDIA Jetson edge compute, ROS 2, PhysX-based simulation, custom reward shaping, and field-operations tooling like TAK/TAKX integration. I write the training code, build the rigs, and drive the robots when it's time to test.

02 / Stack

Tools I reach for every day.

Reinforcement Learning

  • Isaac Lab / Isaac Sim
  • RSL-RL, PPO, GRPO
  • NeMo Gym
  • OpenAI Gym, Gymnasium
  • Reward shaping & curriculum learning

Machine Learning

  • PyTorch, TRL, PEFT
  • HuggingFace Transformers
  • vLLM inference
  • CUDA / Triton
  • LLM agent training

Robotics & Control

  • Boston Dynamics Spot SDK
  • Unitree Go1 control
  • ROS 2 (Humble) / Zenoh
  • Forward kinematics & Jacobians
  • Linear & nonlinear control

Edge & Hardware

  • NVIDIA Jetson AGX Thor / Orin
  • NVIDIA DGX Spark
  • Ubuntu 22.04 / 24.04 (ARM64)
  • Docker, NVIDIA runtime
  • CUDA 13, sm_110a / Blackwell

Simulation & Physics

  • NVIDIA PhysX
  • USD scene authoring
  • Lagrangian / Coriolis dynamics
  • Ballistic & aerodynamic models
  • Domain randomization

Field & Ops

  • TAK / TAKX integration
  • OpenCV, image pipelines
  • Real-time TCP comms
  • Sensor fusion
  • UAV / drone operations
03 / Hardware

The rigs I build on.

Jetson AGX Thor
Edge Compute

Jetson AGX Thor

128 GB unified memory developer kit running native Ubuntu 24.04 ARM64. Used for on-device LLM agent training with NeMo Gym + GRPO and high-throughput inference experiments on Blackwell-class silicon.

  • Memory128 GB unified
  • OSUbuntu 24.04 ARM64
  • CUDA13.0 / sm_110a
  • StackNeMo Gym, vLLM, PEFT
NVIDIA DGX Spark
Workstation

DGX Spark

NVIDIA's desktop AI workstation — the local sandbox for fine-tuning and distillation runs that are too heavy for the edge but don't need a full datacenter. Pairs cleanly with the Thor for asymmetric workflows.

  • ClassPersonal AI Workstation
  • UseTraining & fine-tuning
  • PipelineHuggingFace, TRL
  • Pairs withThor / Orin
AGX Orin mounted on Unitree Go1
Field Robot

AGX Orin × Unitree Go1

A Jetson AGX Orin mounted directly on a Unitree Go1 quadruped — onboard perception, ROS 2 nodes, and policy inference running at the edge with no cloud dependency. The whole platform walks where it's told.

  • ComputeJetson AGX Orin
  • RobotUnitree Go1
  • MiddlewareROS 2 Humble + Zenoh
  • ModeOnboard / untethered
04 / Field Experience

From simulation to the real thing.

Boston Dynamics Spot
Boston Dynamics

Spot — Arm throwing & field ops.

Designed an Isaac Lab RL environment for the Spot arm to perform dynamic ball-throwing — full physics pipeline including forward kinematics, Jacobian-based tip velocity, ballistic trajectory prediction, and a two-ball kinematic / dynamic release system. Currently iterating in version V29 with curriculum learning across distance and lateral targeting.

On the deployment side: zero-shot policy transfer onto real Spot via the Boston Dynamics SDK, including grasp pipelines that fuse fisheye camera data with TAK/TAKX situational awareness for field operations.

SIM// Isaac Lab + PhysX ALGO// PPO via RSL-RL STATE// V29 in training
UAV / Drone
Aerial

Drones & aerial autonomy.

UAV operations work covering flight control, onboard perception, and integration with field tooling. Treats the drone the same way as any other robot: a closed-loop system between policy, sensors, and the physical world.

STACK// ROS 2 FOCUS// Perception + control
05 / Games

In rotation.

Four browser games written in HTML5 Canvas. Side projects keep the fundamentals sharp — and one of them is a thinly-veiled excuse to play with the same ballistic math that drives the Spot throwing environment.

MINECRAFT_VOXEL.HTML
READY
// NOTE: Ballistic shares its core math — gravity-driven trajectories, tip-velocity release, target prediction — with the Spot arm throwing environment. Same physics, different controller. Sometimes the fastest way to debug a reward function is to play with the dynamics yourself.
06 / Projects

Selected open-source work.

Multi-Agent RL

AlphaZero — TicTacToe

AlphaZero implementation with Monte-Carlo Tree Search and joint policy/critic training. Walks through the loss function and self-play loop.

PyTorch · MCTS · Self-play
View on GitHub
Foundations

Neural Networks in PyTorch

Reference implementations of core neural network architectures and training loops — a clean refresher for production work.

PyTorch · CNN · RNN
View on GitHub
Policy Search

CartPole — Hill Climbing

Hill-climbing with adaptive noise scaling on CartPole-v1. Stochastic policy search without gradient information.

Gym · NumPy · Policy search
View on GitHub
Policy Search

MountainCar — Cross-Entropy

Cross-entropy method on MountainCarContinuous. Population-based optimization for continuous control.

Gym · CEM · Continuous control
View on GitHub
Repository Index

More on GitHub →

Full archive of experiments, robotics tooling, and side projects. The visible tip of an iceberg of training runs and lab notebooks.

github.com/bmaxdk
Open Profile
07 / Education

Six courses that shaped the work.

// University of Washington

Graduate coursework — Mechanical Engineering, CSE, Applied Math & EE.

A cross-departmental graduate program covering control theory, optimization, machine learning, computer vision, and parallel computing. The mix that lets me treat robotics as one stack instead of three disconnected fields.

UW Graduate Studies
CSE 571
// 01

AI-Based Mobile Robotics.

The core class for everything I do now. Probabilistic state estimation, motion planning, and decision-making under uncertainty for mobile platforms. Set the foundation for thinking about Spot, Go1, and drones as estimation + control problems first, learning second.

APPLIED IN: Spot · Go1 · Drone Ops
CSE 546
// 02

Machine Learning.

Foundations of supervised, unsupervised, and probabilistic learning — bias / variance, regularization, kernel methods. Everything downstream in deep RL and policy learning rests on these mechanics.

APPLIED IN: RL Pipelines · LLM Fine-tuning
ME 547
// 03

Linear Systems.

State-space modeling, controllability, observability, and the linear-algebra backbone of modern control. This is how you talk to a robot before you start training a policy on top of it. Direct line to the Jacobian-based tip-velocity work in the Spot throwing environment.

APPLIED IN: Spot Arm · Manipulator Dynamics
ME 583
// 04

Nonlinear Control.

Lyapunov methods, sliding-mode and feedback-linearization techniques, stability analysis for nonlinear plants. Real robots are nonlinear. Intuition for stability is what tells you when an RL policy is going to fail before it does.

APPLIED IN: Sim-to-Real · Reward Shaping
CSE 578 / ME 578
// 05

Convex Optimization.

Convex analysis, duality, and algorithms — the mathematical engine room of modern ML and control. Every gradient descent step in PPO is a child of this course. Same for QP-based whole-body controllers and trajectory optimizers.

APPLIED IN: PPO / GRPO · Trajectory Opt
CSE 576
// 06

Computer Vision.

Geometry, multi-view, and learning-based vision — from camera models and calibration through CNNs and detection. Direct application: the fisheye camera grasp pipeline on Spot, and the perception loop on the Go1.

APPLIED IN: Spot Grasp · Go1 Perception

// Additional UW coursework

AMATH 584Applied Linear Algebra & Numerical Analysis ME 574Applied Parallel Computing CSE 573Artificial Intelligence I EE 586Digital Video Coding Systems ME 588Dynamics & Vibrations CSE 415Intro to Artificial Intelligence ME 548Linear Multivariable Control ME 564Mechanical Engineering Analysis I ME 565Mechanical Engineering Analysis II