VLA Gallery / All models / Industry models

Genie Envisioner

Genie Envisioner (GE-Base/GE-Act/GE-Sim)

AgiBot(Genie Team)·LV-NUS Lab·BUAA · 2025-08

VLAIndustry modelsGE-Act 160Mvideo world model

A lightweight action decoder (GE-Act) takes representations from a video world model (GE-Base) to produce actions, and GE-Sim acts as an action-conditioned video simulator.

Paper

Genie Envisioner: A Unified World Foundation Platform for Robotic ManipulationarXiv 2508.05635

Overview

Category

Industry models

Type

VLA

Released

2025-08

Organization

AgiBot(Genie Team)·LV-NUS Lab·BUAA

Open weights

Yes

Parameters

GE-Act 160M; GE-Base based on LTX-Video 2B · COSMOS2 2B (final total not disclosed)

Architecture

Backbone LLM

None (T5-XXL text encoder)

Vision encoder

—

Action head

GE-Act autoregressive action decoder 160M (video world model)

Chunk

54 steps

Control rate

30 Hz

Architecture (original summary)

GE-Base: instruction-conditioned (frozen T5-XXL encoder) multi-view video diffusion DiT (based on LTX-Video 2B · COSMOS2 2B, 5Hz) + GE-Act: 160M autoregressive action decoder (30Hz, 54 steps)

Training

Pretraining accelerators

NVIDIA A100 (GE-Base 32, GE-Act 16)

Pretraining duration

GE-Base Stage I about 7 days + Stage II about 3 days (32 A100 each); GE-Act about 3 days (16 A100)

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

8,832 (Converted · A100-hours (32×168 h + 32×72 h + 16×72 h))

Fine-tuning

A100 ×8, about 12 h (video adaptation) + about 36 h (action specialization)

Data type

Real-robot teleoperation

Data scale

AgiBot-World-Beta about 1M episodes, 2,967 h (introduction wording: about 3,000 h of video-language pairs); 1 h of teleoperation for a new embodiment

Main datasets

AgiBot-World-Beta

Tasks & evaluation

AgiBot G1 5 tasks (sandwich, pouring tea, wiping table, microwave, detergent packing; compared with π0 · UniVLA · GR00T N1); sparse memory analysis E2E 89 / SR 76 (without memory 81 / 64); beats π0 · GO-1 on 3 of 4 RoboTwin tasks; adapted to Dual Franka cloth folding with 1 h of data

Robot

AgiBot G1 (main); adapted to Agilex Cobot Magic and Dual Franka; RoboTwin simulator

Inference

Inference device & speed

Action forward pass of 54 steps (30Hz) takes 200ms on onboard RTX 4090; asynchronous inference with video DiT at 5Hz·actions at 30Hz

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Onboard (RTX 4090)

Reinforcement learning & base

Reinforcement learning

Not used · Video diffusion pretraining + supervised action decoder

Base control

Fixed · No base control (limited to upper-body tabletop manipulation; paper explicitly excludes whole-body locomotion)

Notes

Key features

A lightweight action decoder (GE-Act) takes representations from a video world model (GE-Base) to produce actions, and GE-Sim acts as an action-conditioned video simulator. Among industry models, it disclosed GPU counts and days per stage in the most detail.

Notes

Parameters (B) is null since the total is not disclosed. Pretraining GPU/TPU-hours is the sum of the three GE-Base and GE-Act pretraining stages (about 7, 3, and 3 days converted to 168, 72, and 72 h). Confirmed as a language-instruction-conditioned model (GE-Base instruction-conditioned, T5-XXL), so classified as VLA (corrects the VA/VLA boundary in the notes; could also be seen as Other since it is based on a video world model). Weights released (corrects 'Unverified' in the notes): GE_base 2025-08-14, GE-Sim (Cosmos2) 2025-10-22, GE-Act (Calvin) 2025-12-18. Technical report released 2025-08-08. The GE-Act 'torque' wording is not used because checks of the original were inconsistent.

Verification

Corrections applied