VLA Gallery / All models / Humanoid VLA

EgoVLA

UIUC·MIT·NVIDIA·UC San Diego · 2025-07

VLAHumanoid VLANVILA-2B + action head 300Mregression

Pretrained on human egocentric video, then fine-tuned on robot demonstrations, transferring the human hand action space to humanoid bimanual manipulation.

Paper

EgoVLA: Learning Vision-Language-Action Models from Egocentric Human VideosarXiv 2507.12440

Overview

Category

Humanoid VLA

Type

VLA

Released

2025-07

Organization

UIUC·MIT·NVIDIA·UC San Diego

Open weights

Unverified

Parameters

NVILA-2B + action head 300M (total not disclosed)

Architecture

Backbone LLM

LLM inside NVILA-2B (not stated)

Vision encoder

NVILA vision encoder

Action head

Transformer action head 300M (wrist + MANO) (regression)

Chunk

30 steps (1 s)

Control rate

30 Hz

Architecture (original summary)

NVILA-2B VLM + 300M Transformer action head (6 encoder layers, hidden 1536); action = wrist pose (3D position + rot6D) + top 15 PCA components of MANO hand joints, 1-second horizon (30 Hz × 30 steps)

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed; 100 robot demonstrations per task

Data type

Human video / simulation·synthetic

Data scale

About 500,000 image-action pairs; 100 robot demonstrations per task

Main datasets

4,000 HOI4D videos, 833 min HOT3D, 166 hours HoloAssist (1/10 subsampled), 2,317 TACO motion sequences

Tasks & evaluation

Ego Humanoid Manipulation Benchmark 12 tasks (7 short-horizon · 5 long-horizon); trained backgrounds short-horizon average 77.78%, long-horizon 45.93%

Robot

Sim Unitree H1 + Inspire hands (NVIDIA Isaac Lab)

Inference

Inference device & speed

30 Hz action horizon; hardware not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Human video pretraining + imitation learning

Base control

Fixed · No base control (upper body only: arm EE IK + hand joints, 36-dim; sim H1 legs not controlled)

Notes

Key features

Pretrained on human egocentric video, then fine-tuned on robot demonstrations, transferring the human hand action space to humanoid bimanual manipulation.

Notes

Training setup deferred to supplementary material and absent from the main text (verification NOT FOUND, supplementary material inaccessible). Parameters (B) is null since total parameters are not disclosed. Data hours is null since there is no overall total.

Verification

Confirmed