VLA GalleryAll modelsGallery

VLA Gallery / All models / Reference baselines (pre-2024)

ACT

Stanford University; UC Berkeley; Meta · 2023-04

VAReference baselines (pre-2024)About 80MCVAE

Low-cost bimanual imitation learning policy that regresses action chunks at once and averages overlapping predictions with temporal ensembling.

View architecture diagram in the galleryPaper ↗

Paper

Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwarearXiv 2304.13705

Overview

Category
Reference baselines (pre-2024)
Type
VA
Released
2023-04
Organization
Stanford University; UC Berkeley; Meta
Open weights
No
Parameters
About 80M

Architecture

Backbone LLM
None
Vision encoder
ResNet18
Action head
Joint-position chunk regression + temporal ensembling (CVAE)
Chunk
100 steps
Control rate
50 Hz
Architecture (original summary)
ResNet18 + CVAE Transformer (encoder 4 layers·decoder 7 layers, hidden 512, FF 3200, 8 heads, 32-dim latent); joint position chunk k=100 regression, exponentially weighted temporal ensembling, 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: 1× RTX 2080 Ti (11GB), about 5 hours; batch 8, lr 1e-5, KL weight 10; epochs Not stated in paper (README: 2000 for sim example, 5000+ recommended for real)
Data type
Real-robot teleoperation; simulation·synthetic
Data scale
Real: 50 demos per task (Thread Velcro 100), episodes 8–14 s (400–700 steps), about 10–20 min per task; sim: 50 demos per task (scripted·human)
Main datasets
In-house ALOHA collection
Tasks & evaluation
Real 6 tasks (Slide Ziploc, Slot Battery, Open Cup, Thread Velcro, Prep Tape, Put On Shoe); sim 2 tasks (Transfer Cube, Bimanual Insertion)
Robot
ALOHA low-cost bimanual (leader-follower teleoperation), system under $20,000

Inference

Inference device & speed
About 0.01 s (RTX 2080 Ti)
Inference memory
Not disclosed
Weights bf16 (calc.)
0.2 GB
Where it runs
Local PC

Reinforcement learning & base

Reinforcement learning
Not used · Imitation learning (CVAE behavior cloning)
Base control
Fixed · No base control (ALOHA fixed bimanual)

Notes

Key features
Low-cost bimanual imitation learning policy that regresses action chunks at once and averages overlapping predictions with temporal ensembling. Learned precise bimanual tasks with sub-$20k ALOHA hardware and 50 demonstrations per task.
Notes
All compute figures are per task. Data length is disclosed only in minutes (about 10–20 min per task), so the hours (h) cell is empty. Code·sim data released, no checkpoints.
Verification
Confirmed

Other models in this category