MachineFi Labs

Models for robots and physical AI.

Two models for embodied AI. Trio-Spark selects one of your allowed actions in a single forward pass. Trio-Lumen turns egocentric and robot video into timestamped subtask annotations.

Replays of real output: Trio-Spark on public DROID episodes, Trio-Lumen on a first-person demo clip.
Trio-SparkMultimodal decision model

Closed-set decisions, full probabilities.

A multimodal decision model for physical AI. Trio-Spark selects one allowed action and returns the full probability distribution over your action set in a single forward pass.

  • Text, one image, or a 2–4-frame sampled video window
  • 2–8 allowed actions with caller-defined IDs
  • One chosen action and all action probabilities; no generated text

Who it's forTeams building agents for cameras and robots

$0.042per million input tokensoutput free
Real Trio-Spark output on public DROID episodes: does the clip match its task label?
Trio-LumenRobot video annotation model

Subtask annotation for robot video.

A robot video annotation model for robot learning data. Trio-Lumen segments egocentric and robot video into subtasks and returns timestamped, open-vocabulary captions describing each hand's actions as JSON.

  • Egocentric and robot video, up to 70 minutes each
  • Open-vocabulary subtask captions with left-, right- and both-hand actions
  • JSON output; up to 2,000 videos and 200 video hours per batch

Who it's forRobot learning and data teams

$5per video hourbilled per minute
event_0001_video_0001.json7 segments · 12.4 s
  1. 0.00–3.60

    hold hanger with right hand, pick up shirt with left hand, place shirt on hanger with both hands

    Lpick up shirtRhold hangerL+Rplace shirt on hanger

  2. 3.60–5.00

    hold hanger with left hand, place shirt on hanger with right hand

    Lhold hangerRplace shirt on hanger

  3. 5.00–7.60

    hold hanger with right hand, pick up shirt with left hand

    Lpick up shirtRhold hanger

  4. 7.60–8.60

    place the blue cloth on the table with the left hand while holding the hanger with the right hand

    Lplace the blue cloth on the tableRholding the hanger

  5. 8.60–10.20

    hang hanger on rack with right hand

    Rhang hanger on rack

  6. 8.65–12.40

    place hanger on rack with right hand, pick up blue shirt with both hands

    Rplace hanger on rackL+Rpick up blue shirt

  7. 10.20–12.00

    pick up blue cloth with both hands

    L+Rpick up blue cloth

Exact excerpt from a Trio-Lumen result; left- and right-hand actions split out from each caption.
How it works

Access both models through the API.

  1. 01
    Send the state

    Text, one image or camera frame, or a 2–4-frame sampled video window.

  2. 02
    List the allowed actions

    Two to eight choices with your own IDs and descriptions.

  3. 03
    Act on the result

    Get one choice and a probability for each. Your code decides when to act, ask for review or wait.

  4. Spark API docs
SPARK_IDEMPOTENCY_KEY=$(uuidgen) # once per new decision; reuse it on retries
curl https://platform.machinefi.com/api/spark/v1/decisions \
  -H "Authorization: Bearer $TRIO_SPARK_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPARK_IDEMPOTENCY_KEY" \
  -d '{"model": "trio-spark",
       "task": "Is this support ticket urgent?",
       "state": {"ticket": "Our checkout is down for every customer."},
       "choices": [
         {"id": "urgent", "description": "Needs a response within the hour"},
         {"id": "normal", "description": "Can wait for the normal queue"}]}'
Backed by
Samsung Next
Goodwater Capital
Hanwha
Xoogler Ventures
Escape Velocity

Building robots or physical AI?Talk to us.