VLA Gallery / All models / VA

Mobile ALOHA

Stanford University · 2024-01

VAVANot disclosedCVAE

Low-cost whole-body teleoperated mobile bimanual system; co-trains with static ALOHA data to learn mobile manipulation from 20–50 demonstrations per task.

Paper

Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body TeleoperationarXiv 2401.02117

Overview

Category

VA

Type

VA

Released

2024-01

Organization

Stanford University

Open weights

No

Parameters

Not disclosed

Architecture

Backbone LLM

None

Vision encoder

ResNet18

Action head

16-dim regression of joints + base velocity (CVAE)

Chunk

45 steps

Control rate

50 Hz

Architecture (original summary)

ACT (pretrained ResNet18, 4 encoder layers · 7 decoder layers, hidden 512, FF 3200, 8 heads); 16-dim actions (14-DoF joint positions + base linear · angular velocity), chunk 45; comparison policies DP chunk 64, VINN chunk 100

Training

Pretraining accelerators

N/A (co-training)

Pretraining duration

N/A

Pretraining steps & batch

N/A

GPU/TPU-hours

Not disclosed

Fine-tuning

Per task: hardware·time·steps Not disclosed; ACT batch 16·lr 2e-5, DP batch 32·lr 1e-4 (DDIM 50 training·10 inference steps)

Data type

Real-robot teleoperation

Data scale

50 demos per task (5 tasks)·20 demos (Cook Shrimp, High Five) + co-training with 825 static ALOHA demos (sample ratio 0.5:0.5)

Main datasets

Self-collected Mobile ALOHA; static ALOHA dataset

Tasks & evaluation

Wipe Wine, Cook Shrimp, Rinse Pan, Use Cabinet, Call Elevator, Push Chairs, High Five (7 tasks)

Robot

Mobile ALOHA (mobile base + bimanual 14-DoF)

Inference

Inference device & speed

Collection·inference on an RTX 3070 Ti (8GB) laptop; latency not disclosed; camera 50 Hz

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Onboard (laptop GPU)

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (co-training with static ALOHA data)

Base control

Wheeled (Direct base velocity/pose output) · Policy output includes base target linear and angular velocity (2 of 16 dims)

Notes

Key features

Low-cost whole-body teleoperated mobile bimanual system; co-trains with static ALOHA data to learn mobile manipulation from 20–50 demonstrations per task. System cost is $32k including onboard power and compute.

Notes

Parameters, training hardware, and time all not disclosed (including README). Main policy is ACT; DP and VINN also evaluated.

Verification

Confirmed