NEWAsk your fleet questions in plain language. Meet Ask →
Pulse · Observability

See how your machines behave. Capture what matters.

Monitor models and devices in production. When confidence drops or behavior looks abnormal, capture camera footage, model traces and hardware signals, so your team has the context to improve models and applications.

!
Task success droppedRobot-002 · VLA v2.5 · Abnormal behavior · 2 min ago
Current71%
Expected95–99%
expected 95–99%
Recording triggered
Event recorded · 1 min ago

Works with robots of every shape and size

  • Arms
  • Humanoids
  • Robot dogs
  • Mobile robots
  • Tractors
  • Delivery bots
  • Drones
  • Forklifts
  • Excavators
  • Cleaners
Baselines

Adaptive baselines, per metric.

Pulse learns what normal looks like for every metric on every device, so it can tell a real problem from noise.

  • Task success, confidence, latency and hardware signals
  • Flags unexpected changes on individual devices
  • Finds patterns across the fleet
  • Detects issues even when software reports no error
Task success · last 50 runs
Robot-002 · VLA v2.5
Monitoring
Task successExpected 95–99%
71%
Model confidenceWithin baseline
Normal
Inference latencyWithin baseline
Normal
Recordings

Record the full context.

Around each event, Pulse captures camera footage, model inputs and outputs, sensor readings, motor responses and application logs. Review what the machine saw, what the model produced and what happened next.

Event recording 3/4
Robot-002 · VLA v2.5
Camera footage
Saved
Model inputs and outputs
Saved
Sensor and motor signals
Saved
Application logs
Saving…
Prioritize

Capture the moments worth reviewing.

Severity is weighted by how far a deviation sits from baseline, when it happens and which metrics are affected. You get a ranked signal, not a flat list of alerts.

Event recordings
Latest captures
Unexpected motor responseRobot-017 · motor and sensor signals
35
Low model confidenceRobot-002 · camera and model trace
20
Task failedRobot-031 · camera and app logs
15
Improve

Review, compare and improve.

Route alerts to the right person by severity, team or source without manual triage. Compare recordings across versions to learn from what happens in the field and build better models and applications.

Compare recordings
24 runs · Pick & place · Warehouse A
Version ASuccess and failed runs v2.4
12 runs
Version BSuccess and failed runs v2.5
12 runs
Capabilities

Learn from the field. Build better models.

Models and machines together

Follow model confidence, task outcomes, latency and device health in one platform.

Any slice of the fleet

View data by device, model version or fleet group.

Ranked, not noisy

Severity weighting turns raw deviations into a ranked signal.

Full event context

Camera, model, sensor, motor and log data around every event.

Smart routing

Send alerts to the right person by severity, team or source.

Version comparisons

Compare recordings between model versions to see what improved.

FAQ

Questions, answered.

Something else? Talk to our team.

What does Pulse monitor?

Task success, model confidence, inference latency and hardware signals, with adaptive baselines per metric. It flags unexpected changes on individual devices and patterns across the fleet.

Can it catch issues when there is no error?

Yes. Pulse detects abnormal behavior even when the software reports no error, by comparing each device against its baseline.

What gets recorded?

Around each event, Pulse captures camera footage, model inputs and outputs, sensor readings, motor responses and application logs, so you can see what the machine saw, what the model produced and what happened next.

From your first machine. To the entire fleet.

See Physicalfleet run a live fleet, or talk to the team about yours.

Free for up to 5 devices · Works with robots of every shape and size