Breaking

'mini-vla' Ships: a Vision-Language-Action Model in ~150 Lines

July 1, 2026 at 13:02 EDT

Robotics
Open Source
Foundation Models

Robotics/AI researcher Keivalya Pandya has released 'mini-vla', an open-source Vision-Language-Action (VLA) model that fuses vision, language and robot state to generate continuous actions in roughly 150 lines of code. Explicitly aimed at learning, it is provided under the MIT license on GitHub.

Robot Learning · Open Source

mini-VLA: A Vision-Language-Action Robot Policy in ~150 Lines

A minimal, hackable template that turns images, language instructions and robot state into continuous actions via a diffusion policy — built to teach VLA, not to chase state-of-the-art.

~150

lines of core code
across 4 main files

333

GitHub stars

forks

MIT

open license

Minimal by design — the "Micrograd" of VLA

~150

mini-VLA core

Large VLA

OpenVLA, π₀ — high barrier

Clarity over scale: the same VLA idea, distilled to something a beginner can read end-to-end.

How it works — three inputs, one action

INPUTS

Vision (image) · Language (instruction) · Robot state

→

FUSE

Encoders + fusion combine modalities into one representation

→

ACTION

Diffusion head generates continuous robot actions

Key specifications

Environment
Meta-World MT1 (push-v3), corner camera

Vision encoder
TinyCNN, or CLIP / SigLIP / ViT

Language
GRU + SimpleTokenizer

Training example
Batch 64 · 50 epochs · CPU/GPU

Valued as an entry point

Accessible way into the VLA field
Community added a flow-matching head
Design write-ups well received

Stated limitations

Not state-of-the-art, capabilities limited
No multi-task, no real-robot deployment
Confined to simulation training & testing

What's next

A larger follow-up, Nemotron-VLA built on NVIDIA models, points toward scaling up the backbone and adding multi-task support.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.