VLA GalleryAll modelsGallery

VLA Gallery / All models / Industry models

GR-Dexter

ByteDance Seed · 2025-12

VLAIndustry models4Bdiffusion

Controls a fixed bimanual setup with two 21-DoF humanoid hands (56-DoF total) using a 4B VLA.

View architecture diagram in the galleryPaper ↗

Paper

GR-Dexter Technical ReportarXiv 2512.24210

Overview

Category
Industry models
Type
VLA
Released
2025-12
Organization
ByteDance Seed
Open weights
Unverified
Parameters
4B

Architecture

Backbone LLM
Qwen2.5 · Qwen2.5-VL (size not disclosed)
Vision encoder
Qwen2.5-VL ViT
Action head
action diffusion transformer (88-dim) (diffusion)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
MoT: pretrained Qwen2.5-VL + action diffusion transformer; 88-dim action per timestep (arm joints, 6D EE pose, 16 active DoF per hand, 3D fingertip positions)

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 + web vision-language + other-robot (cross-embodiment) data + human video (including 3D hand·finger tracking)
Data scale
Teleoperation: cosmetics organizing about 20 h, pick-and-place about 20 h; over 800 h of egocentric video; Fourier ActionNet about 140 h; OpenLoong Baihu 100k+ traj; RoboMIND 107k traj (479 tasks)
Main datasets
Fourier ActionNet, OpenLoong Baihu, RoboMIND, GR-3 web VL data, in-house teleoperation
Tasks & evaluation
Long-horizon cosmetics organizing in-domain 0.97 / OOD 0.89; pick-and-place seen 0.93, unseen objects 0.85, unseen instructions 0.83
Robot
2 × Franka Research 3 + ByteDexter V2 21-DoF hands (56-DoF system, fixed bimanual)

Inference

Inference device & speed
Not disclosed
Inference memory
Not disclosed
Weights bf16 (calc.)
8 GB
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Next-token prediction + flow matching joint supervised learning
RL evidence
https://arxiv.org/html/2512.24210
Base control
Fixed · No base control (Franka fixed bimanual + hands)

Notes

Key features
Controls a fixed bimanual setup with two 21-DoF humanoid hands (56-DoF total) using a 4B VLA. About 20 h of teleoperation data per task is supplemented with other-robot data and over 800 h of egocentric video with hand tracking.
Notes
Data hours are stated only per source with no total, so the numeric cell is left empty. arXiv v1 2025-12-30, v2 2026-01-09 (the 2026-07-27 date in the HTML header does not match the submission history). No public repository found (could not be confirmed in verification).
Verification
Confirmed

Other models in this category