Ask & plan
A user gives a goal. A local model interprets it; PRI builds or retrieves an explicit executable task.
PRI · Procedural Robotics Intelligence
AI intelligence is abundant.
Reliable robot execution is not.
The Robotics Execution Harness between the AI brain and the robot body.
Adora Robotics builds PRI — an execution system that turns an AI model’s understanding and plans into physical actions, monitors what actually happens, recovers from changes, and learns missing actions through demonstration. Model proposes. Runtime owns the task.
The missing layer
A brain can decide what to do. A robot still has to make it happen.
An AI model may understand an instruction and propose a plan. Real-world execution still requires object grounding, robot-specific actions, preconditions, completion checks, error handling, and continuous feedback. PRI is the bridge that makes those responsibilities explicit — not merely an SOP generator and not just an app that calls a model.
Architecture
Model proposes. Runtime owns the task.
The model interprets intent. PRI maintains the executable task and supervises its progress. Robot-specific adapters, sensors, and controllers carry out physical action.
REASONING
AI brain
Understand requests, reason about goals, propose or revise a task plan.
EXECUTION HARNESS
PRI runtime
Owns the structured task, selects and sequences atomic actions, evaluates conditions, and handles interruption and recovery.
- Explicit steps, preconditions & completion checks
- Task state, branching & recovery
- Skill library + learned missing actions
- Perception feedback & closed-loop control
EMBODIMENT
Robot body
Sensors, motion control, grippers and hardware-specific action adapters.
Models are replaceable. Physical execution remains governed by PRI. Porting to a different robot requires compatible hardware interfaces and robot-specific adapters.
Execution lifecycle
Ask. Execute. Adapt. Learn when necessary.
Execute & verify
PRI dispatches atomic actions and checks progress using perception and control feedback.
Adapt & recover
If an object moves, an action fails, or the user changes the goal, PRI adjusts the live task and continues.
Learn what is missing
When a required action is unavailable, PRI can request a human demonstration and incorporate a reusable action.
Not a fixed script. The task representation guides execution, but the robot does not blindly replay a recorded trajectory. Feedback, decisions, and recovery remain active throughout execution.
Why this architecture matters
Three separations that make robots more useful.
Reasoning ≠ execution
A stronger general model improves understanding; it does not remove the need for grounded control, task state, and verification.
Procedure ≠ trajectory
A procedure specifies goals, conditions and actions; perception and servo control resolve motion in the current scene.
Task ≠ hardware
Reusable task logic is kept separate from robot-specific action parameters and adapters.
On-robot demonstrations
Tested on real manipulation tasks.
Examples from our current research prototypes — not a claim of unrestricted general-purpose autonomy or production readiness.
Fruit sorting and placement
Pouring beans into a moving target
Cake decorating with mid-task instructions
Hamburger assembly with changes during execution
Shelf pick-and-place and drawer organization
Parameterized milk-tea preparation
Completing a partially finished task
Learning a missing covering action from demonstration
Robot demonstrations are currently focused on tabletop manipulation. New video footage will be published as it becomes available.
Deployment
Designed to work where the robot works.
Our current prototype runs its language and vision models together with the PRI runtime locally on an NVIDIA Jetson AGX Thor. During closed-loop execution, visual feedback can drive corrections without repeatedly consulting the model. This architecture reduces dependence on network connectivity while keeping task execution under local control.
Our direction
Give any capable robot more useful work — without training a new policy for every job.
We are building a model-agnostic execution harness with reusable task memory, new-action acquisition, and robot-specific adapters. Broader hardware support and more demanding tasks remain ongoing development work.