VLA GalleryAll modelsGallery

VLA Gallery / All models / New in 2026

Helix 02

Figure AI · 2026-01

VLANew in 2026S0 10MNot disclosed

S1 expanded from upper body to whole-body control, and S0 is a learned whole-body controller replacing 109,504 lines of hand-coded C++.

View architecture diagram in the galleryOfficial page ↗

Paper

No paper

Overview

Category
New in 2026
Type
VLA
Released
2026-01
Organization
Figure AI
Open weights
No
Parameters
S0 10M; S1·S2 not disclosed

Architecture

Backbone LLM
Not disclosed
Vision encoder
—
Action head
S0 10M learned whole-body controller (Not disclosed)
Chunk
Not disclosed
Control rate
1 kHz (S0) · 200 Hz (S1)
Architecture (original summary)
3 tiers: S2 (slow reasoning·language)·S1 (System 2 latent-conditioned Transformer, whole-body joint targets 200 Hz)·S0 (10M neural network, 1 kHz); action chunk not disclosed

Training

Pretraining accelerators
Not disclosed (S0 uses 200,000+ parallel simulation environments)
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed
Data type
Motion capture (joint-level retargeted human motion, S0) + simulation·synthetic; S1·S2 Not disclosed
Data scale
S0: 1,000+ hours of human motion + 200,000+ parallel environments; S1·S2 not disclosed
Main datasets
Not disclosed
Tasks & evaluation
4 min of continuous autonomous dishwasher unloading·loading (61 locomotion·manipulation motions); bottle cap opening·pill extraction·5 ml syringe (secondary); no standard benchmark
Robot
Figure 03 (secondary)

Inference

Inference device & speed
S1 200 Hz, S0 1 kHz; onboard compute hardware not disclosed
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Onboard

Reinforcement learning & base

Reinforcement learning
Partial (RL low-level controller) · S0 (10M) trained with sim-to-real RL on over 1,000 hours of human motion in 200k parallel sim environments. No RL mentioned for S1·S2
RL evidence
https://www.figure.ai/news/helix-02
Base control
Humanoid legged (Commands/latents → RL locomotion controller) · S1 outputs whole-body joint targets including legs → S0 learned whole-body controller (1 kHz) handles balance·contact; room-scale locomotion + manipulation

Notes

Key features
S1 expanded from upper body to whole-body control, and S0 is a learned whole-body controller replacing 109,504 lines of hand-coded C++. It takes fingertip tactile sensing (3 g detection) and palm cameras as input.
Notes
Release date Figure 2026-01-27, secondary 2026-01-28. Data hours (h) for S0 training only (1,000+ hours).
Verification
Confirmed

Other models in this category