DemoGen-Sim
A simulation reimplementation of DemoGen that turns sparse demonstrations into spatially augmented point-cloud trajectories, with oracle filtering and controlled baselines.
From human videos to robot actions.
I’m Henry Bao, a Duke ECE student building robot learning systems that turn human videos and demonstrations into manipulation policies through imitation learning, reinforcement learning, and simulation.
Selected Projects
A collection of robotics and embodied AI projects focused on how intelligent agents perceive, plan, and interact with the physical world.
Webcam Gesture-Controlled Robotics
A real-time monocular hand-gesture interface that converts webcam landmarks and pseudo-depth cues into stable, safety-aware commands for a simulated Franka robot.
Robot Demonstration: Cloth Folding
A learning-from-demonstration system for dexterous manipulation of deformable objects.
LLM Planners for Embodied Agents
A parallel LLM-based symbolic planning system that drafts multiple action sequences and validates them with an executability checker.
RoboMemory
A full-stack robotics lab assistant that converts robot experiment logs, metrics, notes, and artifacts into persistent agent memory.
Hypernetworks for Implicit Neural Representations
A hypernetwork generating coordinate-based neural representations for images, enabling continuous-resolution querying and high-quality reconstruction.
ROS2 Autonomous Driving
A full-stack autonomous navigation system using ROS2, featuring SLAM, sensor fusion, and MPC-based trajectory tracking.
AI Performance Analytics Platform
A distributed profiling system for monitoring LLM inference latency and throughput across heterogeneous GPU clusters.
AI-Agent GPU Kernel Generation Framework
An LLM-driven compiler that automatically generates and optimizes Triton and CUDA kernels for custom agentic workflows.
Agentic AI Accelerator Design
Research into specialized hardware architectures optimized for the branching and memory-intensive nature of agentic reasoning.
Mini Post-Training Lab for Reasoning Models
A lightweight framework for fine-tuning small language models on chain-of-thought reasoning tasks using RLHF and DPO.
Experience
Duke Robotics / Dexter Lab
Reconstructing object meshes from visual observations, retargeting human hand–object trajectories to robot manipulators, and developing real–sim–real deployment pipelines for manipulation experiments.
Duke University — ECE 383
Supporting students in robotics labs covering ROS2, motion planning, perception, manipulation, and autonomous systems.
Duke University Hospital
Designed and iterated on an electromechanical device to automate breast-biopsy sample handling, achieving approximately 92% reliability in testing.
Technical Skills
Languages, robotics platforms, machine-learning frameworks, and engineering tools.