TwinHands-AI: Bimanual Physical AI

TwinHands-AI is a multimodal Physical AI system designed to transform natural-language instructions into coordinated bimanual robotic actions. The project uses two SO-101 robotic arms in a MuJoCo simulation environment and combines multimodal AI reasoning, computer vision, Vision-Language-Action (VLA) concepts, task planning, motion control, and verification into a single workflow. A user can provide a high-level instruction such as “prepare the dinner table for four people.” Instead of manually controlling the robot, TwinHands-AI interprets the instruction, analyzes the simulated scene, identifies the required objects, generates a task sequence, assigns actions between the left and right arms, executes the manipulation sequence, and verifies the resulting environment. The primary demonstration focuses on autonomous dinner-table preparation for different group sizes. The robots can reason about objects such as plates, bowls, cups, spoons, glasses, and a water jug and coordinate their movements to arrange the table. The system is designed around a Vision-Language-Action pipeline: Natural-language command → Scene understanding → Task planning → Bimanual coordination → Robotic execution → Verification and recovery. TwinHands-AI also explores optimized AI inference using Intel OpenVINO, with an architecture intended to provide a pathway from simulation toward real-world robotic deployment. The broader goal is to demonstrate how modern AI infrastructure can connect generative and multimodal AI with physical-world automation, enabling robots to understand human intent and perform multi-step manipulation tasks autonomously.