VIDOG: VIDRAFT's On-Device AI Kit That Upgrades Remote-Controlled Quadrupeds to Autonomous Walkers
TL;DR: VIDRAFT's VIDOG is a hardware-software AI kit that bolts onto existing low-cost quadruped robots — no firmware overhaul required — and replaces manual remote control with fully on-device autonomous navigation. It fuses a Vision-Language Model (VLM), LiDAR, and a 3D depth camera on an NVIDIA Jetson compute module to deliver edge-local perception and path planning. Developers and robotics integrators working on facility inspection, security patrol, or search-and-rescue applications now have a retrofit path to physical AI without buying new purpose-built autonomous robots.
What it is
VIDOG (비독) is an on-device AI retrofit kit developed by VIDRAFT (비드래프트, CEO Kim Min-sik) that converts commodity remote-controlled quadruped robots into autonomous-walking robots. Rather than selling a complete autonomous robot, VIDRAFT ships the intelligence as a modular unit that sits on top of an existing robot body.
Key characteristics:
- Form factor: A compact AI module mounted on the robot's back — no major modifications to the base chassis or existing firmware are required.
- Target hardware: Currently demonstrated on affordable/entry-level quadruped platforms; roadmap covers wheeled mobile robots and industrial equipment.
- First public deployment: Partner company Sizzle (씨즐) exhibited a VIDOG-equipped autonomous quadruped at the 2026 Gwangju Future Industry Expo, which opened on September 2, 2026.
- Prior validation: VIDRAFT previously deployed an on-device VLM on Boston Dynamics' Spot for Korean-language voice interaction, running as a permanent installation at the Seoul Robot & AI Science Museum. VIDOG extends that same technology to lower-cost hardware.
How it works
VIDOG's architecture is built around three principles: on-device compute, sensor fusion, and model composability.
Hardware stack (conceptual):
- An NVIDIA Jetson-based compute module handles all inference locally — nothing is sent to the cloud during operation.
- A LiDAR sensor provides spatial mapping and obstacle awareness.
- A 3D depth camera supplements visual understanding of the surrounding environment.
- A Vision-Language Model (VLM) processes visual input and, combined with the spatial data, drives perception and decision-making.
Autonomy loop:
- The operator specifies a destination (via interface or, increasingly, Korean-language voice command).
- VIDOG's on-device stack fuses LiDAR + depth + VLM outputs to build a local map and compute a navigation path.
- The robot walks to the target without remote human input.
- On arrival, the robot performs a lie-down gesture as a task-completion signal — a simple but observable feedback mechanism.
Why fully on-device?
Running all inference at the edge eliminates round-trip latency and keeps field data local. This is especially relevant for environments with unreliable or no connectivity — exactly the conditions found in disaster response, underground facilities, and remote inspection sites.
Broader VIDRAFT technology integration:
VIDOG is designed to compose with VIDRAFT's broader ecosystem: their POCKET on-device AI framework and VKAE/VKUE acceleration technologies. The company's stated roadmap layers in local LLM, retrieval-augmented generation (RAG), and voice capabilities so that the robot can eventually answer domain-specific questions about its environment in addition to navigating it.
The business model is explicitly framed as a "Physical AI Foundry" — analogous to an AI model foundry that fine-tunes models for customers, VIDRAFT supplies the process that makes existing robots smarter, remaining hardware-agnostic across robot manufacturers.
Benchmarks & results
No quantitative benchmark figures (latency numbers, navigation success rates, obstacle avoidance accuracy, etc.) were published in the source article. What the article does confirm qualitatively:
- The system has been publicly validated with general visitors at the Seoul Robot & AI Science Museum via the Spot-based on-device VLM deployment — real-world, non-lab exposure.
- VIDOG was live on the expo floor at the 2026 Gwangju Future Industry Expo through partner Sizzle, providing a public demonstration of autonomous navigation on a consumer-grade quadruped.
- The kit operates independently in low-connectivity environments, which is the primary engineering claim differentiating it from cloud-dependent approaches.
Formal benchmark publications have not been announced at the time of writing.
How to try it
VIDOG is currently available through direct B2B engagement with VIDRAFT — there is no public Hugging Face repository, GitHub release, or self-serve API documented in the source article. The kit is being supplied to partner companies (Sizzle being the first named publicly).
If you are a robotics integrator or OEM interested in evaluating VIDOG for your platform, the appropriate channel is to contact VIDRAFT directly. No public endpoints, SDK packages, or model weights have been announced.
FAQ
Q: Does VIDOG require replacing the robot's existing firmware or control software?
A: No. The kit is designed to mount on top of an existing robot without major modifications to the base chassis or firmware. The AI compute module operates as an additive layer.
Q: Which robot platforms is VIDOG compatible with?
A: The first public demonstration used an affordable quadruped platform via partner Sizzle. VIDRAFT has previously worked with Boston Dynamics Spot. The company states its goal is hardware-agnostic compatibility, with plans to extend beyond quadrupeds to wheeled mobile robots and industrial equipment.
Q: Why on-device instead of cloud inference?
A: On-device inference eliminates network latency, enables operation in connectivity-constrained environments (basements, disaster zones, remote sites), and minimizes the exposure of sensitive field data to external networks — all critical for the target use cases of facility inspection, security patrol, and disaster response.
Q: Is Korean-language voice command support currently live?
A: The source article describes Korean voice command and Q&A response as a feature under active expansion — it is not described as fully shipped in the current VIDOG release.
Originally reported by 로봇신문 (2026-09-01) — source article.
Top comments (0)