Dobry Todorov

Overview

Featured at the NVIDIA GTC Taipei 2026 keynote and showcased at Computex 2026.

Built on NeoCore's markerless motion capture technology, NeoCoreRT enables real-time teleoperation of humanoid robots with sub-500 ms end-to-end latency. The system captures full-body and finger motion, retargets it to a remote humanoid robot, and supports sim-to-real deployment through physics simulation for responsive and natural robot control.

The system combines markerless motion capture, real-time retargeting, physics simulation, and networked robot control into a unified low-latency teleoperation pipeline.

Key Features:

• Markerless tracking across configurable capture volumes (3x3 m to 10x10 m)
• Up to 10 synchronized RGB cameras (1080p @ 60 FPS)
• Real-time full-body + finger reconstruction
• <500 ms end-to-end latency on a single PC (RTX 3090 Ti class GPU)
• RReal-time motion retargeting with sim-to-real deployment to humanoid robots

My Role

• Ported NeoCore from Windows to Linux
• Designed the multi-stage GPU pipeline for capture, inference, retargeting, and rendering
• Designed a modular runtime plugin architecture supporting multiple robotics platforms

Tech Stack

C++, Python, CUDA, TensorRT, OpenCV, Mujoco, Isaac Sim

Overview

Featured at the NVIDIA GTC Taipei 2025 keynote and showcased at GTC 2026.

Multi-actor markerless motion capture system using synchronized RGB cameras, delivering full-body and finger tracking at 60 FPS with <200 ms end-to-end latency on a single workstation (RTX 3090 Ti class GPU).

Key Features:

• Multi-actor tracking (1-6 actors) with dynamic capture volumes (3x3 m to 10x10 m)
• Up to 10 synchronized RGB cameras (1080p @ 60 FPS)
• Real-time full-body + finger reconstruction
• <200 ms end-to-end latency on a single PC (RTX 3090 Ti class GPU)
• Real-time streaming to Unreal Engine, Unity, and MotionBuilder

My Role

• Led end-to-end system architecture across capture, inference, IK reconstruction, and streaming
• Designed a multi-camera pipeline handling up to 10x1080p@60FPS streams on a single CPU/GPU system
• Resolved bandwidth and synchronization bottlenecks for multi-camera real-time processing
• Developed a hybrid IK solver reconstructing a 67-joint skeleton from sparse COCO-style keypoints
• Achieved sub-centimeter end-effector accuracy with stable spine and clavicle reconstruction despite missing keypoints
• Optimized GPU inference and skeletal compression to maintain 60 FPS tracking
• Built real-time streaming integrations for Unreal Engine, Unity, and MotionBuilder
• Implemented AWS-based licensing, activation, and feature control backend

Tech Stack

C++, C#, JavaScript, CUDA, TensorRT, OpenCV, Unreal Engine, Unity, WebRTC, AWS

Overview

Co-developed by MSI and Red Pill Lab, DigiME is a real-time avatar system for video calls and streaming, enabling users to animate 3D avatars using a webcam or interact with an AI assistant.

Key Features:

• Webcam-based avatar animation
• Virtual camera integration (Zoom, OBS, Teams)
• AI assistant with 3D interface

My Role

• Led development of a real-time avatar desktop application used in production
• Designed animation and rendering pipeline for webcam-based avatar control
• Implemented virtual camera integration compatible with Zoom, OBS, and Teams
• Integrated LLM, speech-to-text, and text-to-speech APIs for AI-driven interaction
• Scaled product to ~18,000 users with ~1,000+ new users/month
• Optimized pipeline performance to operate within webcam frame rate constraints

Tech Stack

C++, C#, JavaScript, ONNX Runtime, Windows APIs

Overview

A real-time 3D animation software powered by a patented IKNet AI algorithm. It enables users to easily create professional looking animations for film, games, previz, virtual production or live streaming driven by six sensors and a microphone.