Orange humanoid robots sorting parcels along a warehouse conveyor line

Understand the behavior of any robot in real time.

Decisions

A robot makes thousands of decisions a shift.

Complexity

Understanding how a robot thinks is complex, iterative and time‑consuming.

Clarity

That's where Momentum comes in.

See how it works

Decisions

A robot makes thousands of decisions a shift.

Complexity

Understanding how a robot thinks is complex, iterative and time‑consuming.

Clarity

That's where Momentum comes in.

See how it works

Every run carries its own evidence. Every failure arrives with its cause attached.

4: SDK entry points: task, decision, controller_switch, fault
1: required dependency: opentelemetry-api
0: changes to your robot's control logic

How it works

An open convention.
An easy to use SDK.
An engine that never forgets.

pick_and_place.py

with momentum.task("pick_place"):
  momentum.decision(
    type="grasp_selection",
    confidence=result.confidence)

Instrumentmomentum-sdk, next to code you already trust

telemetry video images decision capture release context

Evidencestructured traces, every source attached

failure detected

Momentum agentexplores the evidence when a run fails

✗ fault detectedrun #4412

Root cause — grasp_selection

Confidence collapsed to 0.62 on a reflective part; verification then passed a misplaced piece.

Fix — gate grasps below 0.75 · re-run verification

evidence: 14 traces cited →

Root cause & fixranked, evidence-backed, in seconds

SDK

Up and running before your coffee's cold.

Add momentum.* calls next to the code you already have. We handle the serialization, the attribute naming, the span lifecycle: you keep total control over what your robot actually does.

pick_and_place.py

import momentum_sdk as momentum

# Root span — the task the robot was asked to perform
with momentum.task("pick_and_place", goal=str(target)):

    with tracer.start_as_current_span("skill.grasp"):
        result = grasp_selector.run(point_cloud)
        momentum.decision(
            type="grasp_selection",
            inputs={"object_id": object_id,
                    "confidence": result.confidence},
            output={"grasp_id": result.id},
        )
        momentum.controller_switch(
            from_mode="position", to_mode="impedance",
            reason="contact_detected",
        )
        arm.execute_grasp(result)

The value curve

Useful in the first hour.
Smarter every month after.

Hour one

Drop in a recording of a failed run

run_4412.mcapno code changes

✗ Diverges from run #4411 at grasp_selection

confidence dropped 0.91 → 0.62
Month one

Your engineers confirm each diagnosis

Root cause: grasp_selection✓ confirmed

Remembered, so the next similar failure is recognized in seconds
Month six

Ask questions across your whole history

“Why do picks fail on line 2?”

14 of 17 failures happened below 220 lux, vs 2 of 41 above.