VLA Gallery / All models / VA

ScaleDP

East China Normal University; Midea Group AI Research Center; Shanghai University; Beijing Innovation Center of Humanoid Robotics · 2024-09†

VAVATi 10M, S 33M, B 130M, L 457M, H 1Bdiffusion

Reports performance gains from scaling a diffusion policy transformer from 10M to 1B.

Paper

Scaling Diffusion Policy in Transformer to 1 Billion Parameters for Robotic ManipulationarXiv 2409.14411

Overview

Category

VA

Type

VA

Released

2024-09†

Organization

East China Normal University; Midea Group AI Research Center; Shanghai University; Beijing Innovation Center of Humanoid Robotics

Open weights

Unverified

Parameters

Ti 10M, S 33M, B 130M, L 457M, H 1B

Architecture

Backbone LLM

None

Vision encoder

Not stated

Action head

DiT-style diffusion Transformer (up to 1B) (diffusion)

Chunk

Not stated

Control rate

Not disclosed

Architecture (original summary)

DiT-style diffusion Transformer (AdaLN observation · timestep conditioning, unmasked self-attention); C-step chunk of 6D actions (position x · y · z, rotation roll · pitch · yaw) (C not stated); observation encoder not stated

Training

Pretraining accelerators

N/A

Pretraining duration

N/A

Pretraining steps & batch

N/A

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed (GPU·time·epochs·steps·batch all absent from paper·project page)

Data type

Simulation·synthetic; real-robot teleoperation

Data scale

MetaWorld 20 demos per task; real 100 trajectories per task (close laptop 40)

Main datasets

MetaWorld; in-house Franka·UR5 data

Tasks & evaluation

MetaWorld 50 tasks (3 seeds); real Franka 4 tasks (Close Laptop, Flip Mug, Stack Cube, Place Tennis), bimanual UR5 3 tasks (tennis ball into bag, sweep trash, bimanual cube stacking); vs DP-T MetaWorld +21.6% (project page), real single-arm +36.25%·bimanual +75% (arXiv)/real 4 tasks +22.5% (project page)

Robot

Franka (2 ZED cameras); UR5 bimanual (3 RealSense)

Inference

Inference device & speed

Not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

2 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (behavior cloning) only

Base control

Fixed · No base control (Franka·UR5 bimanual)

Notes

Key features

Reports performance gains from scaling a diffusion policy transformer from 10M to 1B. Despite being a 1B model, it discloses no training compute at all.

Notes

Release month 2024-09 matches the arXiv ID prefix, but the v1 date could not be confirmed (†). The actual improvement differs between the arXiv abstract and the project page, so both are given (correction applied). The project page states ICRA 2025 acceptance; code links to DexVLA GitHub; weights not found.

Verification

Corrections applied