VLA Gallery / All models / Humanoid VLA

Figure Helix

Figure AI · 2025-02

VLAHumanoid VLAS2 7B + S1 80Mregression

Projects the S2 latent vector into the S1 token space, concatenates it with visual features, and trains a single set of weights end-to-end with a regression loss.

Paper

No paper

Overview

Category

Humanoid VLA

Type

VLA

Released

2025-02

Organization

Figure AI

Open weights

No

Parameters

S2 7B + S1 80M (logistics update added +50% to decoder head, new total not disclosed)

Architecture

Backbone LLM

Undisclosed 7B open VLM

Vision encoder

VLM built-in + convolutional vision backbone (S1)

Action head

S1 80M encoder-decoder Transformer (regression)

Chunk

Not disclosed

Control rate

200 Hz

Architecture (original summary)

S2: 7B open VLM (7~9 Hz) → latent vector / S1: 80M cross-attention encoder-decoder Transformer + fully convolutional multiscale vision backbone, 35-DoF continuous actions 200 Hz

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed

Data type

Real-robot teleoperation / human video

Data scale

~500 hours of teleoperation (multiple robots and operators, VLM auto-labeling); logistics scaling study 10~60 hours; Go-Big human video scale not disclosed

Main datasets

In-house teleoperation dataset (private); Project Go-Big egocentric human video

Tasks & evaluation

Picking small household objects, drawer and fridge manipulation, multi-robot handover; logistics 4.05 s per package (from ~5.0 s), barcode success ~95% (from ~70%); going from 10 to 60 hours of demonstrations, handling time ~6.84 s→4.31 s, barcode 88.2%→94.4%

Robot

Figure 02 upper body (wrists · fingers · torso · head), 35-DoF

Inference

Inference device & speed

Model-parallel deployment on 2 low-power embedded GPUs per robot; S2 7~9 Hz, S1 200 Hz

Inference memory

Not disclosed

Weights bf16 (calc.)

14.2 GB

Where it runs

Onboard (2 embedded GPUs)

Reinforcement learning & base

Reinforcement learning

Partial (RL low-level controller) · S2·S1 trained by supervised learning with regression loss. 2025 locomotion handled by a separate neural network trained with sim RL

Base control

Fixed · Upper-body control only (wrists·fingers·torso·head, S1 200 Hz); no mention of legs·locomotion

Notes

Key features

Projects the S2 latent vector into the S1 token space, concatenates it with visual features, and trains a single set of weights end-to-end with a regression loss. The laundry·dishwasher updates only added data, without model·hyperparameter changes.

Notes

Parameters (B) is the sum of 7B + 80M. The S1 vision backbone is initialized from sim pretraining, and S1/S2 inputs are offset in time by the deployment latency. The logistics update (2025-06-07) added Vision Memory, state history, and force feedback. Go-Big (2025-09-18) learned language-instructed navigation from 100% egocentric human video. 2025 locomotion is a separate RL controller (separate row). Helix 02 (2026-01) is out of scope.

Verification

Confirmed