Skip to content

Plan4ARI

Architecture

A two-IPC architecture for industrial robot intelligence

Plan4ARI separates adaptive reasoning from deterministic execution so robot workflows can be planned, checked, monitored and controlled with clear responsibilities.

Two-IPC architecture

Two layers, one control loop

Adaptive reasoning runs on the Open IPC; deterministic, safety-critical execution runs on the Core IPC. Only verified commands cross from one to the other, and the robot state flows back for monitoring and replanning.

Diagram of the Plan4ARI two-IPC architecture. The operator instructs the Open IPC, which runs natural-language and low-code input, the knowledge base, verified PDDL planning, the Behavior Tree and STNU task model and motion planning, supported by sensing and a digital twin. Verified commands go to the Core IPC, where deterministic safety rules, MPC micro-interpolation, neural motion control and the robot and drive interface drive the industrial robot. State, safety and progress return to the digital twin. Trustworthiness spans both layers.
Numbers follow the technical workflow below; M1 to M4 mark the research module responsible for each function.

Open IPC

Cognition, world modelling and planning: user interaction, natural-language interpretation, the knowledge base, symbolic and motion planning, sensing and the digital twin. This layer can evolve quickly while remaining legible to operators and engineers.

Core IPC

Deterministic execution: safety rules, MPC micro-interpolation, neural motion control and the robot and drive interface, all under real-time constraints. It turns verified decisions into reliable industrial robot behaviour.

Technical workflow

From an operator instruction to robot motion in six steps.

Feedback

State, safety and execution progress are returned to the planning layer. Scene updates support replanning when task assumptions change.

  1. Natural-language and low-code input Operators describe the task in natural language or through the MI.RA/Dexter low-code interface.
  2. Knowledge base The request is grounded in a shared model of the cell, the objects and the available skills.
  3. Verified PDDL Symbolic planning produces a task plan that is checked before it is accepted.
  4. Behavior Tree and STNU representation The plan becomes an executable Behavior Tree, with temporal constraints expressed as an STNU.
  5. Motion planning Offline and online motion planning turns each task step into feasible, collision-aware motions, using the latest scene information.
  6. MPC and neural control In the Core IPC, MPC micro-interpolation and neural motion control execute the motion under deterministic safety rules.

Research modules in the architecture

M1

Low-code programming and natural communication

Transforms operator intent into verifiable task representations for industrial robot workflows.

Open module

M2

Intelligent sensing and safe motion planning

Connects scene understanding, feasibility checks and motion planning for adaptive execution.

Open module

M3

Neural motion control

Explores learned control and adaptation while preserving industrial safety constraints.

Open module

M4

Trustworthiness and human-centric optimisation

Frames evaluation around explainability, acceptance, safety and operator confidence across both layers.

Open module

Trustworthiness by design

Safety, transparency, data protection, human supervision and human-centred assessment are treated as design requirements throughout the architecture. This wording describes project design intent and does not claim legal certification.