Explicit authority
Every important decision, state transition, and physical action should have a clear owner. Learning does not grant unrestricted control.
Open source Pre-1.0 Under active development
Build robots that improve without giving up control.
Phoxal is an open-source robotics framework for turning robot experience into observable, repeatable, and deliberately evaluated improvements.
Author the system. Validate it. Simulate it. Execute it. Understand what happened. Prove the next candidate before it becomes the next release.
Control is not the opposite of improvement. It is what makes improvement trustworthy.
The premise
A robot should not become harder to understand as it becomes more capable.
Robotics development can break into disconnected stages. A system is designed one way, simulated another way, deployed through another set of tools, and diagnosed from incomplete evidence after something goes wrong.
Learning can make that problem worse. A new model or policy may appear better, but without repeatable evaluation, explicit promotion, and rollback, improvement becomes hidden change.
Phoxal is being designed around a different premise: the path from an authored robot to its next evaluated release should remain one coherent, inspectable lifecycle.
The North Star
Every execution should produce evidence for a better next release.
Improvement is a controlled loop. Each stage leaves enough evidence for the next decision, and no candidate crosses into operation by accident.
-
Author
Describe the robot, its capabilities, and its intended behavior explicitly.
-
Validate
Reject invalid, incomplete, and incompatible systems before they are allowed to run.
-
Simulate
Exercise the system repeatably without creating a second conceptual model of the robot.
-
Execute
Run with explicit responsibilities, bounded authority, and visible lifecycle transitions.
-
Observe
Capture what happened, when it happened, which version produced it, and which part of the system owned the outcome.
-
Create
Use experience and feedback to produce a candidate model, policy, configuration, or implementation.
-
Evaluate
Test the candidate in deterministic scenarios against explicit acceptance criteria. Reject it when the evidence does not meet them.
-
Promotion boundary
Promote
Accept a proven candidate as the next release, with evidence and a reversible deployment path.
Promote Execute the next release
A promoted candidate becomes the next release, which returns to execution and produces new experience.
Trustworthy improvement
Improvement must remain governable.
These principles make the direction credible. They are boundaries for how a capable robot changes, not optional features to add later.
Deterministic validation
Invalid definitions and incompatible systems should fail before execution, not emerge as ambiguous runtime behavior.
One lifecycle across simulation and reality
Simulation should exercise the same concepts and responsibilities used on hardware, so its evidence remains meaningful.
Observable and attributable execution
Every run should retain enough state, timing, identity, and provenance to understand what happened and why.
Evaluated and reversible change
New behavior should enter operation through explicit evaluation and promotion, with a known version, evidence, fallback, and rollback path.
Development
The North Star is defined. The path is under construction.
Open source Pre-1.0 Under active development
The architecture and implementation may change substantially as we search for the simplest system that can fulfill this lifecycle without compromising its principles.
The project is being built in public. We welcome thoughtful discussion from robotics engineers, researchers, developers, and teams working on reliable autonomous systems.
Built in the open