Physical AI & Robotics · decisions
Decision log
8 decisions, newest first. Each carries its reason and the condition under which it is reversed. JSON: /api/cluster/decisions.
D-008Abstraction only under the four tests· affects stack
- Decision
- An abstraction is written only if it passes at least two of: multiple deployments need it, it reduces integration time, it reduces vendor dependence, it creates a defensible platform capability. Safety authorisation is the one mandatory abstraction.
- Reason
- Abstraction has a cost; robot-agnostic must be earned by a second driver, not declared.
- Reversed when
- Never; the threshold may be raised.
D-007igrimaldi.engineering is the control engine of this cluster· affects this site
- Decision
- This host publishes the registers of the Physical AI cluster: repository registry, architecture, KPIs, decisions, kill register, research program, weekly report and cross-cluster contracts — as pages and as JSON.
- Reason
- The mandate requires a repository registry, architecture map, decision log, kill list, roadmap, customer evidence, competitive intelligence, research backlog and weekly KPI report. One place, machine-readable, is cheaper than nine documents.
- Reversed when
- The cluster is frozen under the kill rule.
D-006Simulation before hardware· affects palletizer-simulation
- Decision
- No hardware is bought or borrowed before the simulator emits the twelve palletizing KPIs on a fixture set.
- Reason
- The cheapest evidence a customer will accept; the mandate forbids hardware-heavy work without a customer hypothesis.
- Reversed when
- A customer supplies the cell and pays for the pilot.
D-005Probabilistic intelligence is separated from deterministic safety· affects stack, palletizer-mcp, palletizer-ros2
- Decision
- No LLM or learned model may command an actuator, modify a safety limit, or bypass the gate. Every physical action passes PLAN → SIMULATE → VALIDATE → AUTHORIZE → EXECUTE → VERIFY.
- Reason
- Safety that can be argued with is not safety.
- Reversed when
- Never.
D-004Autonomous inspection is the bridge to Energy, not a product yet· affects autonomous-inspection
- Decision
- Electrical-infrastructure inspection is developed as research until the seven-step chain (detection → report) is demonstrated end to end.
- Reason
- Premature commercialisation of work near HV assets is the highest-risk move in the cluster.
- Reversed when
- The chain is demonstrated with evidence and a utility or inspection company names a budget.
D-003robot-lidar-fusion is shared IP, benchmarked, not sold· affects robot-lidar-fusion
- Decision
- LiDAR–camera fusion is the perception layer for every cell and robot in the cluster. It is measured on benchmark datasets and never marketed on its own.
- Reason
- Shared across palletizing and inspection; reduces vendor dependence; passes three of the four abstraction tests.
- Reversed when
- A customer asks to buy perception alone with a budget attached.
D-002FloorForge, PaintForge and DryForge remain in Operations· affects boundaries
- Decision
- The three Forge applications are not robotics products of this cluster.
- Reason
- None has physical robotic execution. Concept renders and simulators do not move a robot.
- Reversed when
- A Forge application executes a physical task on real hardware under the safety gate.
D-001Palletizer is the first commercial candidate· affects palletizer
- Decision
- Robot-agnostic palletizing software is the first product target. No other robot vertical is pursued until it has a customer.
- Reason
- Repetitive task, expensive labour, structured environment, understood failure cost, customers with budget, and the only repository in the cluster with a public package and a test suite.
- Reversed when
- Customer discovery produces evidence that a different task scores higher on labour cost × frequency × structure × budget.
Kill register
A robotics project is frozen when it has none of the following. Sunk cost is irrelevant. No new robot vertical without commercial evidence; no hardware-heavy project without a customer hypothesis.
| Subject | A customer | A measurable task | Unique IP | A benchmark advantage | An integration advantage | A plausible ROI | State | Decision by |
|---|---|---|---|---|---|---|---|---|
K-001 palletizer/constructionSecond vertical before the first has a customer; its own changelog calls the figures illustrative. Overlaps Operations. | ✗ | ✗ | ✓ | ✗ | ✓ | ✗ | REVIEW | |
K-002 ai-agent-control (robotics-related parts)Public repository holds a one-line README and archives. No robotics module located. | ✗ | ✗ | ✗ | ✗ | ✗ | ✓ | REVIEW |
✗ = criterion currently missing. FROZEN or ARCHIVED requires a decision above.