Innopolis Robotics Society

Student robotics lab @ Innopolis University ROS 2 · Computer Vision · LLM Agents · Control Theory · Autonomous Vehicles

Innopolis Robotics Society (IRoS) is a student community and lab (Room 105a) at Innopolis University where students build robotics together — from coursework projects to independent R&D efforts involving multi-agent systems, LLMs and computer vision. Some repositories are academic course projects (with sprints, Definition of Done and quality gates), others are open R&D work by individual lab members. Check the Repositories tab above for the always up-to-date list.

The largest and most actively developed repositories in the lab.

LLM-driven multi-robot swarm coordination on ROS 2 Humble. 20 differential-drive robots run simultaneously in a 2D Stage simulation (warehouse / cave / among-us maps), with two MAPF planners (a custom PBS with gradient inflation, plus LNS2), a BehaviorTree.CPP mode dispatcher, leader-follower formations, dynamic obstacles, and a three-channel LLM advisory layer — reactive decisions via /llm/decision, a proactive observer watching Behavior Tree state, and an operator chat with plan preview grounded in a read-only MCP context snapshot of the swarm. Every LLM decision is logged to JSONL for future SFT. The most mature and most-starred repository in the org, with published docs.

IRoS's base mobile robot platform on ROS 2: high-level controller, SLAM (slam_toolbox), Nav2-based path planning, Docker-first launch with a simulator, and hooks for real hardware (Sllidar).

Docker-first ROS 2 Humble workspace for the lab's UR10e manipulator: multiple USB cameras + AprilTags + YOLO object detection + Russian voice control (Vosk) + MoveIt + a SCHUNK EGP-64 gripper. A FastAPI REST gateway sits on top of the ROS 2 stack, so perception and motion can be triggered from a browser, a frontend app, or an LLM/tool-calling agent.

ROS 2 stack for a Unitree A1 quadruped with an onboard Jetson Orin Nano: real-time human pose detection (YOLO-pose, ONNX Runtime), skeleton-based gesture recognition (ST-GCN), and high-level locomotion commands. Runs both on the real robot and locally on a GPU machine for model debugging.

Vision-based Kinematic Imitation — a Video-to-Kinematics pipeline: RGB-D video of a human demonstration (RealSense D435i + Azure Kinect DK) → 3D skeleton extraction (MediaPipe + depth fusion) → retargeting to robot kinematics via object-relative IK (PINK/Pinocchio) → a LeRobot dataset for training ACT / Diffusion Policy. The idea: replace teleoperation with cheap human video without losing trajectory quality.

A hardware + software platform for studying Control Theory, RL and Hardware-in-the-Loop experiments: cart-pole (in progress), double pendulum and "butterfly" (planned), with Arduino firmware and the inno_control Python library.

One physical project spread across several repositories at different stages — from spec to coursework to capstone firmware.

A course project — face recognition + access control for the university lab — built as three independent implementations by three different teams: FaceGuardV1, FaceGuardV2, and FaceGuardV3.

• day_of_innopolis_2026 — a demo for the university's Open Day: a manipulator "writes" text within a given area, using spline generation from text and motion through control points.
• CampPromoBot2025Final — a Python (uv) promo bot for the 2025 summer camp.

• 3D-printing — lab rules for the 3D printers (K1C, K1 Max): model preparation, responsible contacts, safety guidelines and slicer presets (RU/EN).
• .github — the community-health repo this org profile is published from.