Physical AI & Robotics · first target

Palletizer CORE

The target is not a demonstration. The target is: a customer can deploy faster and operate cheaper than with the incumbent alternative. Until a pilot with a kill date says so, every number below is a dash.

github.com/iceccarelli/palletizer pypi · palletizer-full-stack 0.2.0 palletizer-app.vercel.app (stated, not verified here)

Last commit 2026-07-31 (live). Apache-2.0. Python, TypeScript, C++. Detail: /registry/palletizer.

What exists

Deterministic orchestrator loop with hazard evaluation, fault detection and telemetry publication.
Mixed-SKU pallet optimizer with a stability number that can be checked; C++ shelf/skyline packer via pybind11 with tested invariants.
Hardware-agnostic RobotInterface and GripperController; robot-arm profiles; WMS CSV ingestion.
MCP stdio JSON-RPC server exposing planning as agent tools (plan and simulate — never actuate).
ROS 2 integration package and browser physics demos (React Three Fiber + Rapier).
Test suite across orchestrator, planning, optimizer, motion, power, MCP server and native invariants; PyPI package palletizer-full-stack 0.2.0.

What does not exist

A customer, a pilot, a deployed cell, or any measured KPI.
A certified vendor driver for any robot brand. “Robot-agnostic” is an interface, not a compatibility list.
A simulation harness that writes the twelve KPIs as a versioned JSON snapshot (the next experiment).

The twelve KPIs

Palletizing KPIs — 0 of 12 measured
KPI	Definition	Measured	Method
Cycles / hour `cycles/h`	Completed pick-place cycles per hour of scheduled run time.	unmeasured	Simulation harness on a fixed fixture set first; cell telemetry (orchestrator loop) on hardware.
Successful picks `count`	Picks that place the case within tolerance without re-grip.	unmeasured	Orchestrator task log; verified by the post-place check.
Failed picks `count`	Picks aborted, dropped or placed out of tolerance.	unmeasured	Fault detector events joined to task ids.
Intervention rate `per 1,000 cycles`	Human interventions per 1,000 cycles.	unmeasured	Operator acknowledgements in telemetry; the number that decides the economics.
Changeover time `min`	Minutes from last case of SKU set A to first case of SKU set B.	unmeasured	Pattern manager timestamps.
Deployment time `days`	Working days from site survey to first production pallet.	unmeasured	Pilot log, signed by the customer.
Robot utilisation `%`	Share of scheduled time the arm is executing a task.	unmeasured	Controller state telemetry.
Gripper utilisation `%`	Share of cycles using each end effector; idle tooling is cost.	unmeasured	Gripper controller events.
SKU complexity `SKUs/order`	Distinct case geometries per order; mixed-SKU share.	unmeasured	WMS CSV ingestion statistics.
Downtime `min/shift`	Unscheduled minutes per shift attributable to the cell.	unmeasured	Fault detector; classified by cause.
Labour savings `h/shift`	Operator hours displaced per shift, at the customer’s loaded rate.	unmeasured	Customer baseline before pilot; never a modelled figure presented as a result.
Payback period `months`	Months until cumulative savings equal deployment cost.	unmeasured	Derived from the eleven KPIs above and the signed pilot invoice. Absent until then.

Snapshot schema: PalletizingKpiSnapshot 0.1.0. JSON: /api/cluster/kpis.

Shared perception IP — robot-lidar-fusion CORE

ROS 2 LiDAR–camera fusion: SE(3) extrinsic projection, pinhole intrinsics, z-buffer occlusion, KITTI calibration loader, diagnostics CLI. Shared perception IP for palletizing and inspection.

pypi · robot-lidar-fusion 0.4.0github repository not public

LiDAR-fusion benchmark metrics — 0 of 6 measured
KPI	Definition	Measured	Method
Precision `%`	Detected objects that are real, on the benchmark set.	unmeasured	Benchmark dataset with ground truth; KITTI calibration loader exists in 0.4.0.
Recall `%`	Real objects detected, on the benchmark set.	unmeasured	Same benchmark run.
Latency `ms`	Wall-clock from sensor frame to fused output, p50 and p99.	unmeasured	Diagnostics CLI on stated hardware; hardware named with the number.
Failure rate `%`	Frames with no valid fused output.	unmeasured	Diagnostics CLI over the benchmark set, including degraded sequences.
Localisation error `cm`	Position error of fused detections against ground truth.	unmeasured	Benchmark with calibrated extrinsics; error under sensor dropout reported separately.
Generalisation `Δ%`	Metric change when the benchmark scene family changes.	unmeasured	Held-out scene families; occlusion and degraded-environment subsets.

Evaluation covers localisation, object detection, tracking, uncertainty, sensor calibration, sensor failure, degraded environments, occlusion, robustness and latency. Hardware is named with every latency figure.

Autonomous inspection RESEARCH

Electrical-infrastructure inspection is the strategic bridge to the Energy cluster and is not commercialised before the chain is demonstrated end to end:

DETECTION
CLASSIFICATION
LOCALISATION
INSPECTION PLAN
SAFE ROBOT EXECUTION
EVIDENCE
REPORT

Customer outcome: less dangerous human work, lower inspection cost, higher inspection frequency, earlier fault detection. Result contract: InspectionResult 0.1.0. No repository exists yet; the register says so.