Alfred Cueva

I am a Master's student in Robotics and AI at Georgia Tech. I am currently a researcher at the LIDAR Lab, supervised by Prof. Ye Zhao and Zhaoyuan Gu, where I work on learning-based control and motion planning for embodied AI. I also collaborate with the RL2 Lab, advised by Prof. Danfei Xu, working closely with Zhenyang Chen. My research spans the full robotics pipeline from algorithm design and high-fidelity simulation to real-world deployment, with a focus on building scalable sim-to-real frameworks for loco-manipulation in unstructured environments.

Previously, I worked at Samsung as a Robotics & ML Software Engineer for 2 years, where I led end-to-end development of autonomous systems, from perception and motion planning to real-world deployment on various robots. I received my Bachelors in Mechanical Engineering from Seoul National University during which I was a Research Intern at Dynamic Robotics System Lab, advised by Prof. Jaeheung Park and Dr. Daegyu Lim focusing on model-based priors guided Reinforcement Learning for legged robots. I also interned at Soft Robotics & Bionics Lab under the guidance of Prof. Yong-Lae Park on soft robotic multi-modal sensing for industrial robots.

During my time in Korea, I co-organized a non-profit organization dedicated to AI education, AI Tech Play, and helped host the first nationwide AI camp focused on autonomous racing competitions for high school students. I am also passionate about building community and was a founding member of ASAPEC, a student-led organization for underrepresented minorities, where I organized mentorship programs and workshops to support students from diverse backgrounds.

I'm currently looking for internships, feel free to reach out!

Fun Fact: I am fluent in 4 languages.

Show another fun fact

Email  /  Resume  /  Linkedin  /  Github  /  X  /  Instagram

Research

Ultimately, I aim to build general-purpose robots that can reason, plan, and act effectively in the real world, contributing meaningfully to society. My current research focuses on three directions:

• Bridging high-level planning with low-level whole-body motor skills to enable robots to perform complex loco-manipulation tasks.
• Enabling long-horizon reasoning and generalization in vision-language models so that robots can understand and execute complex tasks across open-ended environments.
• Developing principled recipes for RL post-training to improve robustness and real-world performance.

A hierarchical framework joint optimizing a diffusion policy (DP) high-level planner and an RL-based low-level loco-manipulation controller to mitigate distribution shift in offline-trained humanoid systems. The DP is fine tuned via PPO-based policy gradients while the controller adapts to the planner's evolving commands. Achieves >90% success on door traversal and object transport tasks, including extreme out-of-distribution scenarios with smooth real world execution.

PyTorch · MuJoCo · IsaacGym · Diffusion Models · Reinforcement Learning · Whole-Body Control · Sim-to-Real

Vision-language models (VLMs) are used to perform principled frame selection from demonstration videos and extract physics-grounded reward signals for RL training. By grounding reward computation in physically meaningful video frames, robots can learn complex behaviors from raw video without manual reward engineering.

PyTorch · VLMs · Reinforcement Learning · Reward Shaping · Video Understanding

We equip a humanoid robot with large-area tactile sensors on the chest and forearms to enable direct contact force measurement during whole-body manipulation. Integrated into the control pipeline alongside an arm stabilization mechanism, tactile feedback improves robustness across contact-rich tasks including object reorientation, lifting bulky objects, and collaborative transport, highlighting tactile perception as a critical modality for reliable real-world physical interaction.

PyTorch · Diffusion Policy · Tactile Sensing · Whole-Body Control · Sim-to-Real

Led development of autonomous drilling robot for construction sites, deploying rule-based computer vision and motion planning for precise surface drilling with ±2mm accuracy. System handles 20kg payloads and operates in GPS-denied cluttered environments, reducing human exposure to hazardous tasks by 80%. Deployed across 5+ Samsung Factory sites.

ROS2 · OpenCV · Motion Planning · Force Control

Led development and deployment of adaptive AMR fleet (300+ robots) for material transport in construction sites. Implemented safety-aware navigation achieving 92% successful delivery rate in dynamic, GPS-denied environments. System handles 200kg payloads with real-time obstacle avoidance and multi-robot coordination.

ROS2 · LiDAR SLAM · Nav2 · Fleet Management · Motion Planning

Developed meta-RL optimization framework combining model-free learning with physics-based actuator models for bipedal locomotion. Achieved 40% energy reduction in simulated bipedal walking while optimizing parallel elastic actuator stiffness parameters. Framework demonstrated successful sim-to-real transfer potential for weak actuation scenarios.

MuJoCo · PyTorch · MAML · Model-Based RL

Current Position

Built end-to-end machine learning infrastructure for robotics policy development, including automated evaluation pipelines to benchmark performance across simulation and real-world rollouts, with integrated metrics collection, success rate tracking, and failure mode analysis to accelerate research iteration.

Developed Docker-based cloud training environments and custom Isaac Lab tasks across diverse robot embodiments, enabling reproducible experiments on distributed compute clusters and reducing iteration time by 2x.

Designed a VR teleoperation system with real-time retargeting to collect expert demonstrations, cutting data collection time by 60% and producing 500+ high-quality trajectories for policy learning.

Implemented a modular sim-to-real control stack for Unitree G1 humanoid robots, supporting scalable deployment and flexible switching between high-level learning-based policies and low-level controllers.

Engineered a real-time inference server for on-robot policy deployment, enabling low-latency action streaming from large neural network policies over RTC for seamless hardware-in-the-loop evaluation.

PyTorch · Isaac Sim/Lab · USD · Docker · VR Systems · MuJoCo · Diffusion Models · PPO/SAC

Experience

Spearheaded end-to-end YOLOv8 perception pipeline for mobile robots in harsh industrial environments, from dataset creation (10K+ images) to on-device optimization and CI/CD integration, achieving 92% detection accuracy with 30ms inference latency across 5+ Samsung sites.

Led development of precision control and visual SLAM-based localization for 7-DOF manipulator in GPS-denied environments, reducing positioning error by 15% and earning $10,000 award in Smart Construction Challenge from Korean Government.

Engineered ROS2 autonomous navigation stack combining RRT* and Hybrid A* planning with real-time obstacle avoidance, achieving 92% successful delivery rate across 300+ robot fleet handling 200kg payloads in dynamic factory environments.

Built Isaac Sim digital twin workflows with domain randomization for synthetic data generation, reducing data collection costs by 40% while improving model generalization through cross-team sim-to-real validation.

PyTorch · ROS2 · Isaac Sim · YOLOv8 · SLAM · Embedded ML · CI/CD

Developed real-time heat anomaly detection system for semiconductor manufacturing equipment using custom ML architecture and GPU-accelerated pipelines, achieving 95% detection accuracy with <100ms latency for proactive maintenance alerts.

Deployed and optimized Segment Anything (SAM) on industrial edge hardware for collision-aware cluttered-bin retrieval, achieving 3x inference speedup (1.2s to 400ms) through quantization and TensorRT acceleration while maintaining 92% segmentation IoU.

Engineered sensor-fusion collision avoidance system for AGVs combining IMU, LiDAR, and camera data, reducing collision incidents by 80% and enabling safe operation at 1.5m/s in dynamic factory environments.

PyTorch · SAM · OpenCV · GPU Optimization · Sensor Fusion · Real-Time Systems

Proposed novel model-free RL sim-to-real framework for energy-efficient bipedal locomotion with weak actuators, combining meta-RL optimization with physics-based actuator models to achieve 19% speed improvement and 22% energy cost reduction.

MuJoCo · PyTorch · MAML · Model-Based RL

Developed software and system integration for capacitive touch-sensing grid as force-control interface for industrial sewing robots, increasing operation speed by 20% and reducing operator training time by 30% through real-time feedback optimization.

Designed sensor-fusion algorithms with filtering and calibration for stable control signals, collaborating cross-functionally on Software-in-the-Loop validation to reduce control latency by 40% (50ms to 30ms) and improve measurement stability by 25%.

Python · C++ · Sensor Fusion · Capacitive Sensing · Real-Time Control

Projects

Implemented transformer-based diffusion policies for long-horizon visuomotor manipulation, focusing on sequential pick-and-place and multi-block assembly. Integrated camera-calibrated ArUco-based 6DoF pose estimation with action diffusion for closed-loop execution on UR10e robot. Achieved 85% success rate on 3-block assembly tasks with 200+ teleoperated expert demonstrations.

PyTorch · LeRobot · ROS · OpenCV · ArUco Markers · UR10e

Designed end-to-end ROS2 autonomous navigation stack integrating LiDAR-based Bug0 obstacle avoidance with vision-based traffic sign recognition using KNN classifier. Deployed on TurtleBot3 platform for autonomous maze navigation with real-time sign detection and adaptive waypoint generation. Achieved 95% sign recognition accuracy across 10+ maze configurations.

ROS · Python · OpenCV · TensorFlow · LiDAR · TurtleBot3

Developed Task-Invariant Agent (TIA) network for multi-task RL, enabling rapid adaptation to new tasks using model dynamics. The architecture integrates a modified DQN policy network, an encoder for latent task representation from experience sequences, and a model predictor for system dynamics. Achieved 3x faster adaptation to new reward functions compared to baseline DQN, demonstrating robust generalization across CartPole task variants.

PyTorch · OpenAI Gym · DQN · Meta-Learning

Implemented Heuristics Integrated Deep RL approach for online 2D bin packing with placement constraints. Trained PPO agent to learn optimal packing strategies that outperform traditional heuristics. Achieved 15% improvement in space utilization over baseline greedy algorithms.

PyTorch · PPO · OpenAI Gym · Combinatorial Optimization

Implemented 10+ humanoid control algorithms including ZMP-based walking pattern generation, Linear Inverted Pendulum Model, preview control, and whole-body operational space control. Developed CoM estimation using complementary filters and capture point-based stabilization for dynamic walking on simulated bipedal robots.

C++ · MATLAB · Whole-Body Control · QP Solvers · Trajectory Optimization

Developed autonomous driving system for RC Car Racing Challenge using LiDAR-only perception for mapless navigation. Implemented behavior cloning with Gaussian Process Regression to learn driving policy from expert demonstrations. Trained end-to-end control policy mapping raw sensor observations to steering and throttle commands. Achieved top-3 finish in class competition with average lap speed of 2.5 m/s while maintaining safe wall clearance of 15cm.

Python · PyTorch · LiDAR · Gaussian Process · Behavior Cloning

Led research estimating soil health from satellite imagery for agricultural policy enforcement. Managed $8K grant and directed field surveys across 50+ sites, collecting 100GB of GIS and satellite data. Engineered 30+ novel features from multi-spectral satellite data and GIS sources, training machine learning models (XGBoost, Random Forest) for regression. Achieved 60% improvement over baseline estimates (R² = 0.78) using ensemble methods.

Python · Scikit-learn · XGBoost · Sentinel-2 · GIS

Developed novel video generation algorithm that synthesizes realistic video sequences from a single input image using sequential structure learning. Integrated optical flow estimation with temporal consistency constraints to eliminate awkward motion artifacts common in frame-by-frame generation approaches.

PyTorch · Optical Flow · CNNs · Video Synthesis

Template from Jon Barron.