VLA GalleryAll modelsGallery

VLA Gallery / All models / VA

Dense Policy

Shanghai Jiao Tong University; Xidian University; Shanghai Innovation Institute · 2025-03

VAVAAbsolute value not disclosedregression

Generates action sequences in logarithmic time by bidirectional dense expansion with an encoder-only transformer.

View architecture diagram in the galleryPaper ↗

Paper

Dense Policy: Bidirectional Autoregressive Learning of ActionsarXiv 2503.13217

Overview

Category
VA
Type
VA
Released
2025-03
Organization
Shanghai Jiao Tong University; Xidian University; Shanghai Innovation Institute
Open weights
No
Parameters
Absolute value not disclosed; +9.19M vs DP, less than half of ACT

Architecture

Backbone LLM
None
Vision encoder
ResNet18 or 3D sparse convolution
Action head
Bidirectional hierarchical expansion (continuous L2) (regression)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
2D ResNet18 (GroupNorm) or 3D sparse convolutional encoder + 4-layer BERT encoder-only architecture; log-time inference that unfolds the action sequence bidirectionally from coarse to fine from a single frame, continuous action L2 loss

Training

Pretraining accelerators
N/A
Pretraining duration
N/A
Pretraining steps & batch
N/A
GPU/TPU-hours
Not disclosed
Fine-tuning
Per task: GPU·batch·time Not disclosed; 1000 epochs (ACT baseline 2000)
Data type
Simulation·synthetic; real-robot teleoperation
Data scale
10 demos per task for Adroit · MetaWorld, 100 for DexArt; 50 demos per real task
Main datasets
Adroit; DexArt; MetaWorld; in-house Flexiv data
Tasks & evaluation
Sim Adroit, DexArt, MetaWorld; real Put Bread into Pot, Open Drawer, Pour Balls, Flower Arrangement
Robot
Flexiv Rizon + Robotiq 2F-85

Inference

Inference device & speed
Speed similar to ACT, about 10× faster than DP; real experiments on i9-10980XE + RTX 2080 Ti
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Local PC

Reinforcement learning & base

Reinforcement learning
Not used · Imitation learning (behavior cloning) only
Base control
Fixed · No base control (Flexiv arm)

Notes

Key features
Generates action sequences in logarithmic time by bidirectional dense expansion with an encoder-only transformer. It has 9.19M more parameters than DP but inference is about 10× faster.
Notes
Absolute parameter counts are not in the text; Figure 9 compares action head parameters.
Verification
Confirmed

Other models in this category