VLA Gallery / All models / Mobile manipulation

ACT-1

Sunday (Sunday Robotics) · 2025-11

VAMobile manipulationNot disclosedNot disclosed

Converts human demonstrations collected with Skill Capture Glove into robot data via Skill Transform, training long-horizon manipulation and 3D-map-conditioned navigation in one end-to-end model without teleoperated robot trajectories.

Paper

No paper

Overview

Category

Mobile manipulation

Type

VA

Released

2025-11

Organization

Sunday (Sunday Robotics)

Open weights

No

Parameters

Not disclosed

Architecture

Backbone LLM

Not disclosed

Vision encoder

Not disclosed

Action head

Not disclosed (variable-granularity actions from manipulation to navigation) (Not disclosed)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Not disclosed; combines long-horizon manipulation and 3D-map-conditioned navigation in one end-to-end model, generating variable-granularity actions from millimeter-scale manipulation to meter-scale movement

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

Human video

Data scale

Human demonstrations wearing Skill Capture Glove, 0 teleoperation trajectories (official); thousands of gloves deployed (official homepage); 500+ households·about 10 million episodes·thousands of hours (secondary)

Main datasets

In-house Skill Capture Glove household data

Tasks & evaluation

Table clearing→dishwasher loading (33 unique types·68 precise interactions in total, 21 objects, 130+ ft traveled); deployed in an unseen Airbnb home; matching and folding socks; operating an espresso machine; success rates not disclosed

Robot

Sunday Memo home robot (arm reaches floor to 7 ft, normal height 4 ft; wheeled base (secondary))

Inference

Inference device & speed

Not disclosed (homepage only says 'onboard AI')

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not disclosed · Training method not stated in official post

Base control

Wheeled (Format not stated) · A single end-to-end model directly generates even 3D-map-conditioned navigation actions (base command format Not disclosed; wheeled is from a secondary source)

Notes

Key features

Converts human demonstrations collected with Skill Capture Glove into robot data via Skill Transform, training long-horizon manipulation and 3D-map-conditioned navigation in one end-to-end model without teleoperated robot trajectories. It autonomously performed a table→dishwasher task requiring 21 objects·68 precise interactions (33 types)·130+ ft of travel, and was deployed in unseen Airbnb homes.

Notes

Released only via official blog (no paper): architecture·parameters·compute·success rates not disclosed. Glove→robot data conversion (Skill Transform) success rate 90%. Whether language instructions are an input is not disclosed (official post states only 3D map conditioning), so type is provisionally VA. Follow-up ACT-2 Preview (2026-07) announced.

Verification

Confirmed