VLA GalleryAll modelsGallery

VLA Gallery / All models / Industry models

RFM-1

Covariant · 2024-03

VLAIndustry models8BNot disclosed

Handles language instruction understanding, in-context learning, and action-conditioned video prediction in a single 8B model.

View architecture diagram in the galleryOfficial page ↗

Paper

No paper

Overview

Category
Industry models
Type
VLA
Released
2024-03
Organization
Covariant
Open weights
No
Parameters
8B

Architecture

Backbone LLM
Not disclosed (8B any-to-any transformer)
Vision encoder
—
Action head
Action tokens + action-conditioned video prediction (Not disclosed)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
multimodal any-to-any sequence Transformer (text·image·video·robot actions·numeric sensors); uses action-conditioned video prediction as a world model; action chunk·frequency not disclosed

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 operation data + internet data (text·images·video)
Data scale
'tens of millions of trajectories' (press release); (secondary) 'a million trajectories every few weeks', dozens of customers in 15 countries
Main datasets
Covariant real-deployment robot data (private)
Tasks & evaluation
Warehouse pick-and-place; quantitative benchmarks Not disclosed; not deployed to customers at announcement
Robot
Warehouse picking robots (suction, etc.)

Inference

Inference device & speed
Not disclosed (world model video generation about 512×512, about 5fps)
Inference memory
Not disclosed
Weights bf16 (calc.)
16 GB
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not disclosed · Training method not disclosed
Base control
Fixed · No base control (warehouse picking arm)

Notes

Key features
Handles language instruction understanding, in-context learning, and action-conditioned video prediction in a single 8B model. Trained on tens of millions of trajectories of operational data from robots deployed in customer warehouses.
Notes
Announced 2024-03-11 (press-release URL path·IEEE Spectrum article date). (secondary, IEEE Spectrum) Data includes still images·video·joint angles·force·suction strength. The blog calls it the 'largest real-world robot production dataset'.
Verification
Confirmed

Other models in this category