VLA Gallery / All models / Industry models

GR-RL

ByteDance Seed · 2025-12

VLAIndustry models5Bflow matching

Filters demonstrations with a learned task progress (Q-value), then after left-right mirror augmentation runs online RL with a latent noise predictor.

Paper

GR-RL: Going Dexterous and Precise for Long-Horizon Robotic ManipulationarXiv 2512.01801

Overview

Category

Industry models

Type

VLA

Released

2025-12

Organization

ByteDance Seed

Open weights

Unverified

Parameters

5B (VLA + critic MoT in total); online RL noise predictor trainable parameters 51.5M

Architecture

Backbone LLM

Qwen2.5-3B · Qwen2.5-VL-3B config

Vision encoder

Qwen2.5-VL ViT

Action head

flow matching DiT + distributional critic (flow matching)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

MoT: VLA (Qwen2.5-VL-3B-Instruct + flow-matching action DiT) + multi-task distributional critic; latent noise predictor for online RL

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed (evaluated with the online RL 500-step checkpoint)

Data type

Real-robot demonstrations + real-robot online RL

Data scale

Demonstration hours·episode count and online RL time not disclosed

Main datasets

In-house shoelace demonstration data

Tasks & evaluation

Shoelace threading 45.7% (GR-3) → 61.6% (filtering) → 72.7% (symmetry augmentation) → 83.3% (online RL)

Robot

ByteMini-v2 bimanual

Inference

Inference device & speed

Not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

10 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Used (VLA RL post-training) · Filters and augments demonstrations with a distributional critic, then real-robot online RL (trains a 51.5M latent noise predictor). Shoelace success 45.7% → 83.3%

RL evidence

https://arxiv.org/html/2512.01801

Base control

Fixed · No mention of base movement (bimanual precision manipulation of shoelace threading)

Notes

Key features

Filters demonstrations with a learned task progress (Q-value), then after left-right mirror augmentation runs online RL with a latent noise predictor. Raised shoelace threading success from GR-3's 45.7% to 83.3%.

Notes

GR-3 is described as 4B, GR-RL as 5B in total including the critic. Seed page date 2025-12-02, arXiv v1 2025-12-01. No public repository found (could not be verified).

Verification

Confirmed