Different agents, different strengths
I use Codex and Claude for different parts of a problem, then get them to challenge and cross-check each other's work instead of trusting whichever answer arrives first.
Physical AI · Perception · Real-time systems
Research depth,
operational impact.
I want to build physical AI that people will actually use, starting with what operators and markets need and combining robust perception, safe state-aware behaviour, efficient inference, and fleet orchestration to make robotics useful, affordable, and ready to scale.
01 / How I think
Technical work with a reason to exist
The research question and the deployment question belong together.
I am interested in model capability, but also in what determines whether that capability becomes reliable, economical, and valuable in a real operation.
Perception
Give the system a grounded model of the world.
Better geometry and persistent state reduce ambiguity for planning, safety, and manipulation, which also reduces the cost of failures caused by incomplete scene understanding.
Efficiency
Treat latency and compute as design inputs.
Batching, routing, selective computation, and memory-aware execution are what make larger models practical under real response-time and hardware constraints.
Operations
Measure the fleet, not only the demo.
Robot utilisation, idle time, failure recovery, and the usefulness of generated data determine whether a technically capable system delivers value at scale.
02 / Responsible AI practice
Augment judgement, preserve accountability
I use AI to challenge my work, not as a single source of truth.
I care a lot about using AI intelligently and responsibly. I use complementary agents, durable project context, explicit verification, and scheduled experiment monitoring because moving faster only matters if the work still holds up.
Evidence before acceptance
I check claims against source files, logs, experiment results, and reproducible tests. If two agents disagree, I want that disagreement surfaced rather than quietly smoothed over.
Context that survives the session
I keep structured project records, handoffs, and status files so decisions don't disappear when a chat ends, which really matters when research runs across weeks or months.
Research that keeps moving
I pair long-running RL training with scheduled agent check-ins that inspect progress, flag failures, and preserve results outside working hours, including weekends, while bringing decisions that need a person back to me.
03 / Current agenda
Work in development
Three layers of dependable physical AI.
These are active directions rather than finished claims: policy robustness, an independent safety layer, and the orchestration needed for large deployments.
01
New project · Sim-to-real efficiency
Generalised RL policies with efficient large-model inference
Developing a simulation-first pipeline that trains robot policies across varied terrain, object textures, and lighting. The generalised model is paired with batching and efficient serving strategies so larger models can operate within practical latency and compute limits.
02
Research direction · Physical AI safety
Persistent safety memory above the task policy
Exploring a sim-first safety layer that retains state about vulnerable objects and constrains unsafe motion, even when an underlying task model proposes a trajectory that could collide with a person, pet, device, or fragile object.
03
Systems concept · Fleet orchestration
Make robot activity legible and useful at fleet scale
Designing an orchestration layer that connects task allocation with analysis of robot-generated data and outputs, so factory deployments can be coordinated and improved as a system rather than as isolated machines.
04 / Selected work
Research translated into systems
Evidence from models, runtimes, and real hardware.
FPGA-Accelerated Edge Vision
Built and deployed PCB-defect classification and ADAS detection on a Xilinx ZCU104, including a custom Verilog 3×3 convolution accelerator, INT8 quantisation, and confidence recalibration.
3.3× faster6.3× lower energy85.3% accuracy44.6% mAP@0.5
Deployment relevance: measured accuracy, latency, recall, and energy together rather than optimising an offline metric in isolation.
LiDAR-Initialized 3D Gaussian Scene Representation
Replacing a camera-based probabilistic initializer with fixed-size LiDAR anchors while preserving the downstream Gaussian representation. Current work is focused on real-data validation and a controlled baseline.
LiDAR priors3D GaussiansControlled comparison
Deployment relevance: stronger geometric priors make the scene model beneath planning and safety more grounded and inspectable.
Geometry before appearanceRepresentative LiDAR point-cloud structure. Reference image only, not a project result.
Daniel L. Lu · CC BY 4.0 ↗Robot Learning, Retargeting & Real-Hardware Infrastructure
Developing and evaluating VLA, imitation-learning, reinforcement-learning, and motion-retargeting workflows across UR5/UR5e, Unitree G1, and RoboParty V1, supported by ROS 2 control paths, simulation infrastructure, calibration, trajectory auditing, and safe teleoperation.
RLMotion retargetingIsaac LabROS 2
Deployment relevance: turns learned motion into a safer, testable path from simulation and demonstrations to real hardware.
Train in simulation, test for realityRepresentative G1 simulation environment from an open-source Unitree project. Not Panasonic footage.
Source · Unitree Robotics ↗GELLO + SpaceMouse Teleoperation
A dual-path teleoperation stack combining open-source GELLO hardware with browser-based SpaceMouse input through WebHID, rosbridge, and MoveIt 2 Servo, including a yaw-only constraint.
WebHIDROS 2MoveIt 2 Servo
Deployment relevance: more precise, safer interfaces support more consistent robot demonstrations and cleaner training data.
05 / Experience
Technical depth + commercial context
Research, deployment, and product judgement.
My technical work is complemented by business-development and product experience, which shapes how I evaluate usefulness, adoption, and scale.
AI & Robotics Research Intern · Panasonic R&D Singapore
VLA evaluation, robot-learning infrastructure, calibration, teleoperation, and real-hardware deployment within the Intelligent Perception & 3D team.
Undergraduate Researcher · NTU EEE
LiDAR initialization for 3D Gaussian scene representations under Prof. Tay Wee Peng, working with Dr. Loo Junn Yong.
Business Development · MLDA@EEE
Industry partnerships and applied-AI initiatives with NVIDIA, Micron, and collaborators.
Product Owner & Go-to-Market · Golden Horizon
Led an AI analytics product, built event-to-price data pipelines, and analysed 20+ competitors to shape positioning and execution.
Undergraduate Researcher · NTU EEE
Deployment-aware edge-vision research with Dr. Loo Xi Sung, culminating in the IEEE ICIEA 2026 paper below.
06 / Publication
IEEE ICIEA 2026 · Catania, Italy
Deployment constraints are part of the research problem.
2026
FPGA-Accelerated Edge Vision: PCB Defect Detection and ADAS Object Detection on Xilinx ZCU104
Completed through URECA Year 2, the work covers classification, detection, custom hardware acceleration, quantisation, and confidence recalibration under measured performance and energy constraints.
A. Prakash and X. S. Loo · Presented at the 21st IEEE Conference on Industrial Electronics and Applications.