VLA Gallery / All models / Industry models

DYNA-1

Dyna Robotics · 2025-06

기타Industry modelsNot disclosedNot disclosed

Finds and fixes its own mistakes with a reward model that evaluates task progress, aiming for long unattended operation in repetitive commercial manipulation.

Paper

No paper

Overview

Category

Industry models

Type

기타

Released

2025-06

Organization

Dyna Robotics

Open weights

No

Parameters

Not disclosed

Architecture

Backbone LLM

Not disclosed

Vision encoder

—

Action head

Not disclosed (Not disclosed)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Not disclosed (RM-in-the-loop training, putting the reward model into the training loop)

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed (2025-11 pretraining post: 100% success cases with about 1 hour of demonstrations — cup stacking, celery slicing, etc.)

Data type

Not disclosed

Data scale

Not disclosed

Main datasets

Not disclosed

Tasks & evaluation

24-hour autonomous napkin folding of over 850 napkins, 99.4% success rate, about 60% of human speed, 0 interventions; laundry folding · cup filling; laundry folding · parcel sorting without post-training (2025-11)

Robot

Fixed-arm commercial manipulation robot (model not disclosed)

Inference

Inference device & speed

Not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Partial (Reward/preference optimization) · RM-in-the-loop training that puts the reward model (RM) into the training loop. RL algorithm not specified

Base control

Fixed · No mention of base control (bimanual napkin·towel folding and other manipulation)

Notes

Key features

Finds and fixes its own mistakes with a reward model that evaluates task progress, aiming for long unattended operation in repetitive commercial manipulation. Reported 850+ napkins folded in 24 hours with a 99.4% success rate.

Notes

(secondary, AI Business 2025-05-06) 99%, over 700 units, 60% of human throughput, model release planned per report — the primary Dyna page (2025-06-24) values take precedence. Research post dates: Dyna-1 2025-06-24, Open-World Dexterity 2025-10-15, Pre-Training 2025-11-20. Architecture not disclosed, so type is Other.

Verification

Confirmed