Industry · 01 / Physical AI

Every episode your robots run, on one timeline.

Joint torque, gripper cameras, depth and transforms from the whole fleet — captured losslessly, searchable in seconds, and ready to train on.

1.2 kHz
joint telemetry in sync with 30 fps video
LeRobot
episodes export ready-to-train
Fleet-wide
every unit, every episode, one record
01 · Phloem for Physical AI

Proprioception and vision, finally in one place.

Phloem keeps every modality of an episode on one playhead, so debugging a grasp takes minutes, not an afternoon of file juggling.

Scrub the whole episode

Torque, current, wrist camera and depth snap to the same instant. See exactly what the policy saw when the grasp slipped.

Find the failure fleet-wide

Query every episode for the same slip signature. One anomaly becomes a labelled set of every occurrence.

Curate training data from real runs

Tag episodes, filter by outcome, and export LeRobot-format datasets straight from the timeline.

02 · Lifecycle

One platform, from first prototype to fleet.

The record built in development carries into validation and operations — nothing starts from scratch.

01

Develop

  • Record episodes at the bench, network or not — they land on one timeline and sync when connected.
  • Compare controller revisions run against run, joint by joint.
  • Keep sim rollouts and real episodes side by side.
02

Validate

  • Run policy evals as repeatable suites, scored on every revision.
  • Codify pass criteria once; every future episode is checked automatically.
  • Build edge-case libraries from real failures for regression.
03

Operate

  • Capture the deployed fleet continuously, at full rate.
  • Trend joint wear and drift across units before they fail.
  • Route anomalies to tickets with the full episode attached.
03 · One timeline

What a run looks like in Phloem.

A fleet episode as it lands: proprioception, wrist camera, depth and transforms, already on one playhead.

Fleet Capture bench · recordingcloud · in sync UNIT 12 / 24 · LIVE
03:12elapsed
842Kmessages
5streams
1.8 GBon disk
T+00:00T+01:00T+02:00T+03:00T+04:00
joint/arm/joint0/pos 0.42 rad
video/cam/front 30 fps
lidar/lidar/points 16k pts
force/arm/gripper/force_n 18 N
The scene, not just signals

Depth, detections and masks sit over the camera frame, with the point cloud beside them — an episode reads as a scene, not a folder of files.

grasp target points · 3d bin · 0.94 /cam/front · boxes2d · seg
04 · Ecosystem

Native to the robot stack.

First-class ingest for the formats, protocols and buses this work already runs on — and the wider ecosystem behind them.

datasetsLeRobot
frameworkROS 2
formatMCAP
simIsaac Sim
trainPyTorch
protocolCAN bus

And anything else with an SDK, a topic, or a recorded log.

05 · Outcomes

What teams get back.

hours → minutes
Episode triage

Overnight fleet episodes arrive pre-correlated, so morning review starts at the anomalies instead of at the file system.

Nightly
Fleet data → training sets

Curated episodes flow from the fleet into training-ready datasets on a schedule, not as a quarterly project.

06 · End of run

Robots that learn from every episode.

Bring the whole fleet’s data into one place, and turn every grasp — landed or missed — into the next policy.