Multistage-VLA
Euler Robotics' Multistage VLA (Multi-stage Vision-Language-Action) is an AI control architecture that lets robots understand their work environment and generate the optimal action.
With simple voice or text commands, users can direct robot motion, control signals and peripheral devices, and assign pick-and-place tasks for a wide variety of objects, while working flexibly alongside the robot in the same space.
A Physical AI–Based Automation System
By applying the right AI model and algorithm to each stage, Multistage-VLA is designed to deliver both
the reliable task execution and the processing speed that industrial sites demand.
This approach enables a Physical AI–based automation system in which robots understand
unstructured physical environments and turn that understanding into real work.
How It Works
To overcome the task success rate and inference speed limitations that single end-to-end VLA models face in industrial settings, we split the robot's decision-making process into multiple stages.
-
1Voice/text input processing
-
2Vision-based scene understanding
-
33D spatial perception
-
4Task planning
-
5Robot execution
Key Features of Multistage-VLA
01Physical AI Ready for Industrial Sites
With fast cycle times and high task success rates, Multistage-VLA is a Physical AI system built for real-world industrial deployment.
02Natural Language Robot Control for Human–Robot Collaboration
Operators communicate with the robot using natural language commands alone and can collaborate flexibly with it in the same workspace.
03Handles Unstructured Environments and Integrates with Your Systems
It understands target objects and task context even in unstructured work environments, and integrates with a wide range of sensors, robots and peripheral systems to extend the scope of on-site automation.
Specifications
Detailed specifications for technical evaluation.
| Architecture | Multistage VLA (multi-stage Vision-Language-Action) |
|---|---|
| Processing stages | Voice/text input → visual perception → 3D spatial perception → task planning → robot execution (5 stages) |
| Language input | Natural language via voice or text (Korean, English) |
| Visual perception | VLM-based object recognition (down to brand, label and SKU level) |
| Spatial perception | Center coordinates, size and pose extraction using a ToF 3D camera |
| Agent architecture | LLM-based agent (decomposes and executes multi-condition instructions) |
| Command handling | Decomposes and executes natural language commands for robot-executable tasks such as grasping, transfer and stacking |
| Robot compatibility | Hyundai Robotics, Kawasaki, Doosan, FANUC and more (manufacturer-agnostic) |
| System integration | RBS palletizing/depalletizing, AMRs, conveyors |
| Competitive advantage | Higher task success rate and faster inference than end-to-end VLA |
Ongoing Collaboration
Multistage-VLA's technology is being validated through government program selection and real-world on-site PoCs.
Selected for 2025 Deep Tech TIPS
Selected for Korea's Ministry of SMEs and Startups' Deep Tech TIPS program, we are receiving approximately KRW 2 billion in R&D funding. Building on Multistage VLA technology, we are developing RBS 2.0 — end-to-end automation from unloading through sorting and stacking.
PoC with LG Science Park
We are jointly running PoCs in four areas, including Multistage-VLA, with LG Science Park. These include SKU picking and transfer based on semantic grasping, and building a testbed for collecting VLA training data.
Looking for the right product or solution?
Request a Demo & Site Assessment