Robotics Software Engineer · Hanoi, Vietnam

PHAM DUC DUY.

Closing the gap between
autonomous navigation research and deployed hardware.

Shipped ROS 2 robots in hospitals, restaurants, and factories. At VinRobotics, I build factory integrations and VLA manipulation systems; at HaiBotLab, I lead Embodied AI research. Co-authored a MethodsX Q2 paper on distributional RL path planning.

Manipulation · locomotion · field robotics

What I actually do.

Portrait of Phạm Đức Duy
Phạm Đức Duy · Hanoi

Most robotics demos work in ideal conditions. My job is to make them work when they don't — when the LiDAR scan is partial, the floor is reflective, the network drops packets, and the operator expects the robot to keep moving. I tune DDS QoS profiles and ROS 2 executor strategies for real-time performance, configure Cartographer SLAM for constrained hardware, and debug Nav2 failure modes that never appear in simulation.

At ByteHome Robotics, I led the full cycle from prototype to commercialized product — hardware integration, software stack, and the rarely-mentioned last mile: writing operator handbooks for independent fleet operation without ongoing engineering support.

At HaiBotLab I'm focused on the harder problem — moving beyond scripted navigation toward robots that understand unstructured 3D environments. We work with depth images, point cloud pipelines, and Micro-ROS on ESP32 running RTOS for low-latency sensor communication.

Core Stack

Robotics
ROS & ROS 2Nav2 CartographerMicro-ROS GazeboMuJoCo
AI / Perception
Embodied AIReinforcement Learning Imitation LearningPoint Cloud
Primary Languages
C++Python
Familiar Languages
RustGo
Tools
LinuxDocker
Hardware
Visionary-TIntel D435 ESP32Raspberry Pi

Where I've built things.

Mar 2026 — Present
Active
VinRobotics
Robotics Software Engineer · VinRobotics Talent Program
  • Designed and implemented a robot license management system — cryptographic activation, feature gating, and telemetry reporting — for fleet deployments.
  • Integrated software for factory automation projects at VF, supporting assembly-line QC workflows and reducing manual inspection steps.
  • Working on VLA for robotic arms on rich-contact tasks, including pick-and-insert.
Oct 2025 — Mar 2026
ByteHome Robotics
Robotics Software Engineer · Tech Lead (5-person team)
  • At a small robotics startup, led a 5-engineer team building differential-drive service robots for hospitality and healthcare — full ROS 2 stack: sensor drivers, Cartographer SLAM, Nav2 planners/controllers, multilingual voice interaction.
  • Took the product from internal prototype to commercial deployment, including hardware integration, safety testing, and regulatory documentation.
  • Wrote operator handbooks for independent fleet operation without ongoing engineering support.
Jun 2024 — Present
Active
HaiBotLab ↗ haibot-lab
Lab Leader · Hanoi Intelligence Robot Laboratory
  • Direct lab research on Embodied AI: representation learning from egocentric observations, manipulation policy training in MuJoCo, sim-to-real transfer of perception modules.
  • Built point cloud and depth image processing pipelines (Intel D435, Visionary-T) for 3D object detection in unstructured environments.
  • Developed Micro-ROS firmware on ESP32 with FreeRTOS — sub-5ms sensor publish latency enabling tight control loops between embedded hardware and the ROS 2 compute graph.
Graduated Jun 2026
Hanoi University of Industry
B.Eng. · Robotics and Artificial Intelligence
  • 2nd Prize, University Scientific Research 2026 — Satellite image map recognition via image processing and deep learning.
  • 2nd Prize, University Informatics Olympiad 2025.
  • 3rd Prize, University Scientific Research 2025 — Object detection from depth sensor data. Consolation Prize 2024 — Pick-and-place trajectory optimization on a TOSHIBA DC650B die-casting cell.
  • 3rd Prize, University Student Mathematics Competition 2023 (Linear Algebra).

Projects I've shipped.

Real robot
Imitation Learning · Thesis Work
HG-Dagger Obstacle Avoidance
131×fewer FLOPs per control cycle vs DWA
Local obstacle avoidance policy trained via Human-Gated DAgger. The learned policy replaces the DWA local planner — a single forward pass on LiDAR scan input yields 131× fewer FLOPs per control cycle versus DWA's iterative trajectory sampling. Deployed on Micro-ROS / ESP32.
Micro-ROSFreeRTOS ROS 2C++
Real robot
Outdoor Mobile Robotics · University Research
Skid-Steer Field Robot
60 kgpayload · outdoor platform
60 kg payload outdoor platform with skid-steer kinematics. Motor controllers communicate via RS-485 Modbus on a Micro-ROS ESP32 bridge, decoupling real-time motor control from main compute. Supports RC teleoperation and autonomous ROS 2 navigation.
Micro-ROSFreeRTOS RS-485ROS 2
Simulation · mjlab
Reinforcement Learning · Manipulation
UR5e Force-Aware Peg-in-Hole
0.5 mmradial clearance · 2,048 parallel envs
Pick-and-insert task for a UR5e with a Robotiq 2F-85 gripper and a wrist force/torque sensor. A state-based PPO policy grasps a randomly placed peg and inserts it into a randomly placed hole with 0.5 mm radial clearance, trained on 2048 parallel environments in MuJoCo Warp (mjlab) with a curriculum that tightens the hole clearance. The environment exposes observations and actions in ForceVLA format for force-aware imitation learning.
PythonPyTorch MuJoCoPPOForce/Torque
Simulation · mjlab
Reinforcement Learning · Sim-to-Real
Microduck Bipedal Locomotion
4,096parallel MuJoCo Warp envs
Velocity-tracking walking policy for Microduck, a 25 cm, 14-servo bipedal robot. Trained with PPO on 4096 parallel MuJoCo Warp environments (mjlab) against a voltage-controlled servo actuator model, with domain randomization for sim-to-real transfer. Exported to ONNX and validated in a CPU MuJoCo rehearsal before deployment on hardware. My work focused on the training and deployment pipeline; the robot platform is from Pollen Robotics.
PythonPyTorch MuJoCoPPOONNX
Paper figure Center simulation figure from the published paper, showing a robot navigating around obstacles.
Reinforcement Learning · Published Research
Path Planning via Distributional RL
Q2journal paper · MethodsX 2025
The research behind the MethodsX Q2 publication. Trained a QR-DQN distributional critic for mobile robot navigation in multi-structured environments, modeling return uncertainty under noisy depth-sensor input.
PythonPyTorch ROSGazebo
Gameplay · ALE
Reinforcement Learning · Foundations
Atari Agent — Double DQN
From scratchDouble DQN in PyTorch
Implemented Double DQN from scratch to understand the overestimation bias in standard DQN and how decoupling action selection from value estimation corrects it — foundational work that informed the distributional RL research above.
PythonPyTorchGymnasium

Published work & recognition.

2026
2nd Prize · University Scientific Research
Satellite Image Map Recognition Based on Image Processing and Deep Learning
Recognition pipeline for satellite map imagery using classical image processing for region segmentation combined with a CNN-based classification backbone.
2025
Q2 Journal MethodsX · Dec 2025 · vol. 15, 103554
Path Planning for Mobile Robots in Multi-Structured Environments via Distributional Reinforcement Learning
Standard RL path planners optimize for expected reward, which makes them brittle when sensor noise is heavy-tailed. This paper replaces the expected-value critic with a QR-DQN distributional critic that models return uncertainty for navigation across structured, semi-structured, and cluttered map variants in Gazebo.
DOI Reference
2nd Prize · Informatics Olympiad
University Informatics Olympiad 2025
Placed 2nd in the competitive programming track.
3rd Prize · University Scientific Research
Object Detection from Depth Image Data
Detection architectures trained on depth images, targeting robustness in low-light and texture-poor environments typical of indoor robot operation.
2024
Consolation Prize · Univ. Scientific Research
Motion Trajectory Design and Optimization for a Pick-and-Place Robot on the TOSHIBA DC650B Die-Casting Machine
Designed and optimized end-effector trajectories for an industrial pick-and-place arm integrated into a die-casting production cell, focusing on cycle time reduction and collision-free path generation.
2023
3rd Prize · Student Mathematics Competition
University Student Mathematics Competition 2023 — Linear Algebra
Placed 3rd in the Linear Algebra category of the university-wide student mathematics competition.

Let's work on something real.

Looking for roles where the robots actually ship — not just simulate. Particularly interested in teams working on autonomous mobile robots, Embodied AI, or edge-compute robotics stacks where the distance between a ROS 2 node and a physical actuator is short.

Research collaborations welcome: sim-to-real transfer, distributional RL, or perception for navigation. If you're training operators on ROS 2 and need someone who's done it before — that too.

Send me an email ↗

Based in Hanoi, Vietnam.
Open to international roles — remote-first or relocation.