VLA Gallery / All models / VA

ALOHA Unleashed

Google DeepMind · 2024-10

VAVA217Mdiffusion

Trains a per-task diffusion transformer on thousands of teleoperation demonstrations per task, achieving precise bimanual tasks such as shoelace tying.

Paper

ALOHA Unleashed: A Simple Recipe for Robot DexterityarXiv 2410.13126

Overview

Category

VA

Type

VA

Released

2024-10

Organization

Google DeepMind

Open weights

Unverified

Parameters

217M (Base), 150M (Small: encoder 17M + decoder 37M)

Architecture

Backbone LLM

None

Vision encoder

ResNet50 × 4 cameras

Action head

Diffusion denoiser decoder (55M) (diffusion)

Chunk

50 steps (1 s)

Control rate

50 Hz

Architecture (original summary)

Per-camera ResNet50 for 4 cameras + transformer encoder (85M) and diffusion denoiser decoder (55M); 50 diffusion training steps, 50-action chunks (1 s), target above 50 Hz

Training

Pretraining accelerators

N/A

Pretraining duration

N/A

Pretraining steps & batch

N/A

GPU/TPU-hours

Not disclosed

Fine-tuning

Per-task models: TPUv5e 64 chips (JAX data parallel), batch 256, 2M steps (about 265 hours); Adam, weight decay 0.001, lr 1e-4 after 5000 warmup steps

Data type

Real-robot teleoperation; simulation·synthetic (sim teleoperation)

Data scale

Shirt 8,658 (Easy 5,345·Messy 3,313), Lace 5,133 (2,212·2,921), FingerReplace 5,247, GearInsert 4,005, RandomKitchen 3,198 (in-domain 216); real total 26k+ episodes; sim 2,000+ (introduction)/522·201·550 (Table 3)

Main datasets

In-house ALOHA 2 collection

Tasks & evaluation

Real: ShirtEasy·ShirtMessy, LaceEasy·LaceMessy, FingerReplace, GearInsert-1·2·3, RandomKitchen (Bowl·+Cup·+Fork); sim: SingleInsertion, DoubleInsertion, MugOnPlate

Robot

ALOHA 2 bimanual (10 robots, 2 buildings)

Inference

Inference device & speed

Forward pass + iterative denoising 0.043 s (RTX 4090)

Inference memory

Not disclosed

Weights bf16 (calc.)

0.4 GB

Where it runs

Local PC

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (diffusion)

Base control

Fixed · No base control (ALOHA 2 fixed bimanual)

Notes

Key features

Trains a per-task diffusion transformer on thousands of teleoperation demonstrations per task, achieving precise bimanual tasks such as shoelace tying. 35 operators collected 26k+ episodes over 8 months with 10 robots.

Notes

The paper does not explain how the introduction's '2,000+' sim demos relate to the per-task values in Table 3. '2,000+' is an introduction sentence, not from the abstract (correction applied). Sim demos were also collected by teleoperation.

Verification

Corrections applied