We research, design, and deploy intelligent systems at the convergence of artificial intelligence, embedded IoT, autonomous robotics, and physical computing. Every prototype we build ships with a research paper.
Our research spans the full stack — from foundation models running on edge silicon to autonomous robotic systems operating in the physical world.
Vision-language-action models, flow matching, world models, and reinforcement learning for robotics. We train, fine-tune, and deploy cutting-edge models.
ESP32-S3, edge ML inference, sensor fusion, MQTT, and micro-ROS. We put intelligence where the data is born — on-device, low-power, real-time.
Autonomous manipulation, visual servoing, TAMP, and foundation models for robotics. Sim-to-real transfer in MuJoCo and Isaac Sim.
Custom PCBs, sensor modules, motor drivers, and embedded firmware. We design, fabricate, and program the hardware that AI runs on.
Every project ships with open-source code, CAD files, schematics, and a detailed research write-up.
A fully autonomous robot that accepts natural language commands, processes them via cloud LLM APIs (Groq/OpenRouter), and executes motor actions in real-time. Built from the ground up with custom PCB, CAD-designed chassis, and ESP-IDF firmware.
Six FreeRTOS tasks running across dual cores with inter-task queues for sensor data, motion commands, and UI state. Micro-ROS integration for ROS2 compatibility. On-device camera with JPEG capture and base64 encoding for vision-language queries.
Active research program investigating the gap between end-to-end VLA models and modular symbolic TAMP approaches for robotic manipulation. Evaluating state-of-the-art architectures (TiPToP, MolmoAct2, GR00T N1.5) across semantic and distraction-heavy tasks.
Published analysis showing modular TAMP still outperforms monolithic VLA on semantic/distractor benchmarks in 2026. Investigating flow matching for trajectory generation and asynchronous reasoning for real-time robotic control.
Production pipeline for deploying quantized neural networks on ESP32-S3 microcontrollers. Supports ONNX → ESP-DL conversion, INT8 quantization, and real-time inference at <1W power draw.
Benchmarked inference: 8 FPS for MobileNetV2-class models, 2 FPS for lightweight transformers. Integrated with MQTT telemetry for fleet monitoring and OTA model updates via WiFi.
Unmol AI Technology is an independent research and development lab operating at the frontier of embodied artificial intelligence. We don't just publish papers — we build the hardware, write the firmware, train the models, and deploy the complete system.
Founded on the belief that AI research must be grounded in physical reality, our team combines expertise in deep learning, embedded systems, mechanical design, and robotics to create fully integrated intelligent systems.
Every project we undertake produces open-source artifacts: CAD models, PCB schematics, firmware source code, training datasets, and comprehensive technical documentation. Our work has been benchmarked against the best in the field — and we share our findings openly.
Collaborate With UsWhether you're interested in research collaboration, custom hardware design, or deploying AI on the edge — we'd love to hear from you.
We are actively seeking collaboration on VLA architectures, edge AI deployment, and sim-to-real transfer. If you have a research problem that needs both hardware and ML expertise, reach out.