Embodied AI & Robotic Vending Kiosks

Beyond Coils: The Case for Robotic Vending

Traditional vending machines have a fundamental limitation: every product must fit neatly into a coil, a tray, or a fixed compartment. Change your product lineup — change the coils. This works fine for standardised snacks and drinks, but it falls apart for irregular products, fragile items, or anything that requires gentle handling.

KioskForce robotic kiosks solve this. Instead of mechanical dispensing, a robotic arm with AI vision picks each product individually from free-form storage bins — just like a human would.

• Any product shape — no coil pitch, no tray reconfiguration. The arm adapts.
• AI vision guidance — the system sees each product, verifies it, and picks the right one.
• Gentle grasping — vacuum gripping handles fragile items that would break in a traditional vend.
• Free-form bin storage — products sit loosely in cooled or ambient bins, packed to maximise density.

KICR: Our First Robotic Vending Prototype

KICR (KioskForce Ice Cream Robot) is our flagship embodied AI vending machine, currently in active development. It demonstrates the full robotic vending pipeline:

• 4-axis self-developed SCARA arm — 600mm reach, <0.2mm positioning accuracy at full extension. Joint-side magnetic encoders (AS5048A) with closed-loop PID control on every axis.
• AI vision system — RGB camera + YOLO object detection identifies products in the freezer bin, computes grasp coordinates, and guides the arm.
• Vacuum gripper with pressure feedback — silicone suction cups with real-time pressure monitoring detect successful grasps and trigger retries on failure.
• Distributed real-time control — 4× STM32F103 microcontrollers over RS-485 (Dynamixel 2.0 protocol) manage each joint independently, with a Raspberry Pi running inverse kinematics and trajectory planning.
• -20°C operation — the entire picking system operates inside a chest freezer, engineered for reliable cold-chain product handling.

Development Timeline

What Embodied AI Means for Vending

Traditional vending is deterministic — the machine executes a fixed sequence: rotate motor N by X degrees, product falls. If the product jams, the machine doesn’t know. If the bin is empty, it can’t adapt.

Embodied AI vending is different. The system:

1. Sees — the camera scans the storage area and builds a real-time understanding of what’s where
2. Plans — the planner computes an optimal grasp trajectory, avoiding obstacles
3. Acts — the arm executes the pick, with pressure sensing confirming success
4. Learns — every pick feeds back into the system, improving grasp strategies over time

This isn’t just about picking ice cream. The same architecture applies to:

• PPE dispensing — pick safety glasses, gloves, respirators from free-form bins without mechanical reconfiguration
• Pharmaceutical retrieval — handle delicate medication packaging with gentle vacuum grasping
• General retail — vend any product shape without coil setup

Technical Foundations

Our robotic platform is built on principles that differentiate it from off-the-shelf industrial arms:

Self-Developed SCARA Arm

Instead of buying a $10,000+ industrial SCARA arm, KioskForce engineered our own — DC motors with timing belt reduction, joint-side AS5048A magnetic encoders for backlash-free closed-loop control, and aluminium links CNC-machined in Hebei. Result: industrial-grade positioning at a fraction of the cost, with full control over the kinematics stack.

Distributed Real-Time Architecture

Each joint is an independent STM32F103 microcontroller running a PID control loop at 1kHz. Joints communicate over RS-485 using the Dynamixel 2.0 protocol — the same protocol used by $50,000+ industrial robots. A Sync Write packet updates all 4 joints simultaneously, enabling smooth coordinated motion. The Raspberry Pi handles the non-real-time layers: inverse kinematics, trajectory planning, AI vision inference, and the user interface.

AI Vision Pipeline

A RGB camera mounted above the storage area feeds into a YOLO object detection model running on the main controller. The model identifies product types, locates graspable surfaces, and outputs world coordinates via a calibrated camera-to-arm transform. The system handles real-world variance — products shifted during previous picks, uneven bin surfaces, and varying lighting conditions.

Beyond Ice Cream: The Robotic Kiosk Platform

KICR is the first application, but the robotic kiosk platform is designed to be general-purpose. The same arm, vision system, and control architecture can be re-purposed for:

Interested in robotic vending for your application? Contact us to discuss how our robotic kiosk platform can be adapted to your product category. We’re actively seeking early-adopter partners for our 2026 pilot programme.