FrozenVLA-2B — Lightweight Vision-Language-Action for Robot Manipulation

A 50M-parameter trainable head on top of frozen Qwen3-VL-2B-Instruct, trained on BridgeData v2 (2,617 episodes) for general-purpose robotic manipulation.

🧠 Architecture

┌──────────────────────────────────────────┐
│         Qwen3-VL-2B-Instruct (FROZEN)    │
│  ┌─────────────┐    ┌─────────────────┐  │
│  │ Vision Enc. │ -> │   LLM (28L)     │  │
│  └─────────────┘    │  hidden=2048    │  │
│                     └────────┬────────┘  │
│                              │           │
│                     last_token_hidden    │
│                         (2048-dim)       │
└──────────────────────────────┼───────────┘
                               │
                    ┌──────────▼──────────┐
                    │   MLP Projector     │  ← Trainable (1.3M)
                    │  2048 → 512 → 512   │
                    │  GELU + Dropout     │
                    └──────────┬──────────┘
                               │
                    ┌──────────▼──────────┐
                    │   Action Head       │  ← Trainable
                    │    512 → 7          │
                    └──────────┬──────────┘
                               │
                    EEF delta: [dx, dy, dz, ax, ay, az, gripper]

Component	Params	Status
Qwen3-VL-2B (Vision + LLM)	2,127,532,032	❄️ Frozen
MLP Projector	1,314,824	🔥 Trained
Action Head	513	🔥 Trained
Total	2,128,847,369	1.3M trainable

📊 Training

Item	Detail
Dataset	BridgeData v2 (LeRobot format)
Episodes	2,617
Action format	EEF delta 7D: [dx, dy, dz, ax, ay, az, gripper]
Action normalization	Z-score (mean/std per dimension)
Hardware	Alibaba Cloud PAI · NVIDIA A10 24GB
Framework	PyTorch 2.6.0 · Transformers 4.51+ · CUDA 12.6
Epochs	4 (of 5 planned)
Effective batch size	64 (32 × gradient_accumulation=2)
Optimizer	AdamW (lr=1e-4, wd=1e-2)
Schedule	Cosine annealing · 500 warmup steps
Precision	bfloat16 · gradient clip=1.0
Image size	448×448 (Qwen3-VL default)
Frames	All frames per episode (frame_sampling=all)

Training Loss (per epoch)

Epoch	Approx MSE Loss	Checkpoint
1	—	`epoch_1.pt` (15MB)
2	—	`epoch_2.pt` (15MB)
3	—	`epoch_3.pt` (15MB)
4	—	`epoch_4.pt` (15MB)

Inference Requirements

Platform	VRAM	Notes
RTX 4060 Laptop 8GB	~4 GB	✅ Verified · sdpa · bf16
A10 24GB	~4 GB	✅ Training env
T4 16GB	~4 GB	✅ Should work

🚀 Quick Start

1. Clone & Install

# Requirements
pip install torch>=2.5.0 transformers>=4.51.0 accelerate sentencepiece protobuf Pillow

# Clone this repo
git clone https://huggingface.co/YOUR_USERNAME/frozenvla
cd frozenvla

2. Download Base Model

# Qwen3-VL-2B-Instruct (from HuggingFace)
huggingface-cli download Qwen/Qwen3-VL-2B-Instruct --local-dir ./Qwen3-VL-2B-Instruct

3. Load & Infer

import torch
from PIL import Image
from model import FrozenVLA

# Load model with trained head
model = FrozenVLA(
    llm_name="./Qwen3-VL-2B-Instruct",
    mlp_hidden_dim=512,
    mlp_depth=2,
    action_dim=7,
    attn_implementation="sdpa",  # "flash_attention_2" on A100/H100
)
model.load_trainable("epoch_4.pt")
model = model.to("cuda").eval()

# Cast head to bf16 (to match Qwen3-VL output dtype)
model.mlp_projector = model.mlp_projector.to(dtype=torch.bfloat16)
model.action_head = model.action_head.to(dtype=torch.bfloat16)

# Inference
image = Image.open("robot_view.jpg").convert("RGB")
instruction = "pick up the red block"

with torch.no_grad():
    action = model([image], [instruction])
    # action: (1, 7) EEF delta [dx, dy, dz, ax, ay, az, gripper]
    print(action.float().cpu().numpy())

4. Deploy Script (One-Click)

python deploy.py --checkpoint epoch_4.pt --test-image robot_view.jpg

📁 Files

File	Description
`epoch_4.pt`	Best trained head (MLP + ActionHead, 15MB)
`epoch_1~3.pt`	Intermediate checkpoints
`model.py`	Full model architecture (FrozenVLA class)
`config.yaml`	Training configuration
`deploy.py`	One-click deployment + inference script

⚠️ Limitations

Action space: BridgeData EEF delta only — NOT directly compatible with joint-space robots without IK conversion.
Domain: Trained on BridgeData scenes (tabletop manipulation). Zero-shot generalization to novel environments is limited.
Single image input: Uses the current frame only; no temporal context from video history.
Language: English instructions only.

📝 Citation

@misc{frozenvla-2026,
  title  = {FrozenVLA-2B: Lightweight VLA from Frozen Qwen3-VL on BridgeData},
  author = {},
  year   = {2026},
  url    = {https://huggingface.co/YOUR_USERNAME/frozenvla}
}

📄 License

Apache 2.0

Downloads last month: 8

Video Preview

Robotics