HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

Figure 03Figure—BotQFigure—HelixFigure—HelixFigure—Figure 03Figure—Figure 03Figure—HelixFigure—Figure—
Helix
Courtesy of Figure AI · Press kit, editorial use (opens figure.ai)

Helix

by FigureUS

PilotStage 3 of 5

Helix 2.5 (Sep 2026) claims to work in 30 homes it had never seen.

Updated 23d agoChecked 25 Sep0 updates this week

Milestones

Next · Reliable autonomy in customer homes
  1. Figure ends OpenAI collaboration to build in-house AI4 Feb 2025Complete.
  2. Helix unveiled20 Feb 2025Complete.
  3. Logistics package sorting at scale7 Jun 2025Complete.
  4. Laundry folding and dishwasher loading demosSep 2025Complete.
  5. Helix 02 with learned whole-body control (S0)27 Jan 2026Complete.
  6. Helix 2.5: zero-shot generalisation across 30 homes17 Sep 2026Complete.
  7. Reliable autonomy in customer homesNowCurrent milestone.

Most important updates

  • 23d ago
  • 3 Sep 2026
  • 27 Jan 2026
  • 20 Feb 2025
  • 4 Feb 2025

Current obstacles

  • Data coverageAny new object or room can cause a failure. Figure is adding human video (Go-Big) and a paid data platform (Index).
  • Onboard compute limitsThe robot's battery and heat limits cap how big and fast its AI can be.

Physics limits

  • There is no internet of robot motionLanguage models learn from trillions of words already online. Robot data must be recorded in real time, one robot at a time, so even the largest sets are tiny by comparison.
  • Errors compound over long tasksIf each step of a task succeeds 95% of the time, a 20-step chore succeeds only about 36% of the time. Long household jobs need near-perfect steps or reliable recovery from mistakes.
  • Onboard compute is bounded by power and heatLarge models are slow; a robot must react to slips within milliseconds. Chips inside the body share the battery and must shed their heat, which is why Helix splits a slow planner from a fast controller.

How it works

3 parts
Helix folding towels on a Figure 02: one neural network turns what the cameras see into finger and arm motions
Helix folding towels on a Figure 02: one neural network turns what the cameras see into finger and arm motionsCourtesy of Figure AI · Press kit, editorial use (opens figure.ai)
S2

Understand the scene

An AI trained on internet images and text reads the cameras and the task a few times a second, and sets a goal.

S1

Fast motion

A small model turns that goal into joint moves 200 times a second, so the robot reacts fast even though planning is slow.

S0

Whole-body controller

Helix 02 adds a layer trained on 1,000+ hours of human motion that keeps balance and coordinates legs, torso and arms.

Spec sheet

SpecHelix
S2 model size~7B parametersR (reported)
S2 rate7-9 HzR (reported)
S1 model size~80M parametersR (reported)
S1 control rate200 HzR (reported)
Action space (Helix 1)Upper body, 35 DoFR (reported)
S0 training data1,000+ hours of human motionR (reported)

R reported by the company

Papers & demos

4 items
  1. 27 Jan 2026demo
    Helix 02: full-body autonomy (opens figure.ai)A 4-minute, 61-step dishwasher task across a kitchen, with walking, grasping and balance in one AI.
  2. 18 Sep 2025blog
    Project Go-Big: internet-scale humanoid pretraining and human-to-robot transfer (opens figure.ai)Learns to navigate from first-person human video, a way around the shortage of robot training data.
  3. 7 Jun 2025demo
    Scaling Helix: a new state of the art in humanoid logistics (opens figure.ai)Package sorting got faster and more accurate with more data, a sign the AI improves with more examples.
  4. 20 Feb 2025blog
    Helix: a vision-language-action model for generalist humanoid control (opens figure.ai)First VLA to run fast control of a full humanoid upper body onboard, including two robots working together.

Update log

6 updates

Thu 17 Sep

  • Major: BlogSoftware

Thu 3 Sep

  • Minor: BlogPartnership

Tue 27 Jan

  • Major: BlogSoftware

Wed 17 Sep 2025

  • Minor: BlogPartnership

Thu 20 Feb 2025

  • Major: BlogSoftware

Tue 4 Feb 2025

  • Major: XPartnership

About Figure

The team behind Helix

Figure

Humanoids, Helix AI, BotQ factory

Figure builds electric humanoid robots and Helix, the AI that controls them. Founded in 2022 by Brett Adcock, its robots work at BMW's Spartanburg plant, and in 2026 its BotQ factory built about 1 robot per hour.

  • Figure 02 worked about 11 months at BMW Spartanburg. Figure 03 arrived there in June 2026.
  • All its AI is in-house: Helix (Feb 2025), Helix 02 (Jan 2026) and Helix 2.5 (Sep 2026).
  • BotQ went from 1 robot a day (Jan 2026) to 1 an hour (Apr 2026). First line: 12,000 robots a year.

figure.ai (opens figure.ai)Company profile →

Founded
20224 yrs
Headquarters
United States
Status
Private
Valuation
$39BprivateSep 2025
Staff
~600Jan 2025
Raised
$1.7B3 rounds
Last round
Series C · $1B16 Sep 2025$39B post
Works in
RoboticsHumanoidsAIManufacturing
Coverage
3 programs · 15 updateslatest 23d agochecked 25 Sep
Lead investors
Parkway Venture Capital
Customers
BMW Group (opens bmwgroup.com)Catalyst Brands (opens catalystbrands.com)
Partners
Brookfield (opens brookfield.com)OpenAI
Investors
NVIDIA
Suppliers
Nscale (opens nscale.com)
People
Brett Adcock (opens en.wikipedia.org)Founder & CEOCorey Lynch (opens coreylynch.github.io)Head of AI (Helix)
Primer