Figure AI

Helix 02

Figure AI, United States · January 2026

Figure AI, Inc.

Foundation VLAClosed weightsAlso written Helix-02, Helix 2, Figure Helix 02
Parameters
not published
Figure publishes only the System 0 size for Helix 02: a 10M-parameter neural network.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

Helix 02 is the January 2026 revision of Figure's humanoid control stack and the current version as of August 2026. It keeps the System 2 reasoner and System 1 visuomotor policy and adds System 0, a 10M-parameter whole-body controller running at 1 kHz that Figure trained purely in simulation on over 1,000 hours of retargeted human motion across more than 200,000 parallel environments. The practical change is that locomotion and manipulation are now one learned system on the Figure 03, rather than a walking controller handing off to a manipulation policy. The public evidence is one four-minute dishwasher run of 61 consecutive actions. As with Helix 01, nothing is released and nothing is benchmarked against other models.

Architecture

Backbone
Three tiers. S2 handles semantic reasoning, scene interpretation and behaviour sequencing. S1 is the visuomotor policy that produces full-body joint targets at 200 Hz from head cameras, palm cameras, fingertip tactile sensors and full-body proprioception. S0 is new: a 10M-parameter whole-body controller executing at 1 kHz for balance, contact and coordination, trained entirely in simulation.
Action head
S1 emits complete joint-level targets for the entire robot at 200 Hz. S0 converts those targets into 1 kHz joint commands, replacing what Figure states was 109,504 lines of hand-engineered C++ whole-body control.
Control rate
200 Hz
Parameters
Figure publishes only the System 0 size for Helix 02: a 10M-parameter neural network. It does not restate S1 or S2 parameter counts for this version. The February 2025 Helix post gave 7B for S2 and 80M for S1, but Figure does not confirm those numbers still apply to Helix 02, so they must not be carried over.
Pretraining data
System 0 was trained entirely in simulation across more than 200,000 parallel environments on over 1,000 hours of joint-level retargeted human motion data. Figure does not state the S1 or S2 training corpus for Helix 02.
Embodiments
Figure 03 humanoid, full body including locomotion, with palm cameras, fingertip tactile sensing that Figure states detects forces as small as three grams, and full-body proprioception

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Published figures. These come from the model authors, not from this site.
Control rate200 Hz

Figure publishes no GPU model, no memory figure and no millisecond latency for Helix 02. The only rate figures are 200 Hz for S1 and 1 kHz for S0.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Figure internal demonstration (kitchen dishwasher) Figure 03, one continuous autonomous run in a real kitchen
    61 actions
    consecutive loco-manipulation actions without human intervention or resetVendor-reported single demonstration, not a benchmark, no success rate over repeated trials and no third-party verification.source
  • Figure internal demonstration (kitchen dishwasher) Figure 03, same run
    4 min
    duration of the continuous end-to-end autonomous runVendor-reported. Figure gives no distribution over repeated attempts.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Figure demonstrated Helix 02 on a Figure 03 in a real kitchen, autonomously unloading and reloading a dishwasher as a four-minute end-to-end run of 61 consecutive loco-manipulation actions without human intervention or system reset. Follow-up posts in 2026 show Helix 02 doing a living-room tidy (March 2026) and a bedroom tidy (May 2026). No success rates, no trial counts and no independent evaluation are published for any of these.

Fine tuning it yourself

Not possible for third parties. No weights, no code and no API are released.

Where it helps, where it does not

Strengths

  • Extends learned control down to the whole-body layer. S0 is a 10M network at 1 kHz that Figure states replaced 109,504 lines of hand-engineered C++ balance and contact code.
  • Removes the walk-then-act split. Locomotion and manipulation are driven by one learned stack rather than separate modules handing off to each other.
  • Each tier runs at its own natural timescale, so a 1 kHz balance loop, a 200 Hz visuomotor policy and a slow semantic reasoner coexist in one system.
  • Adds fingertip tactile sensing and palm cameras as policy inputs, with Figure stating force sensitivity down to three grams.

Limits

  • Fully closed. No weights, no code, no paper, no API. Every figure comes from one Figure blog post.
  • The only performance evidence is a single four-minute dishwasher demonstration with 61 sequential actions. There is no success rate across repeated trials, no task suite and no public benchmark score.
  • Figure publishes a parameter count only for S0 (10M). S1 and S2 sizes for this version are not disclosed, so the actual compute footprint cannot be estimated.
  • S0 was trained entirely in simulation, and Figure publishes no sim-to-real gap measurement.
  • Bound to the Figure 03 platform, including its tactile and palm-camera hardware. Nothing here is portable to another robot.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.