ilessio-aiflowlab/project_loki

LOKI — HDINO-T Open-Vocabulary Detector

Part of the ANIMA Perception Suite by Robot Flow Labs.

Paper

HDINO: A Concise and Efficient Open-Vocabulary Detector (Mar 2026) Hao Zhang, Yiqun Wang, Qinran Lin, Runze Fan, Yong Li Paper | Code

Architecture

• Backbone: Swin-T (96-dim, depths 2-2-6-2)
• Transformer: 6 encoder + 6 decoder layers, deformable attention
• Text encoder: CLIP ViT-B/32 (frozen)
• Detection head: contrastive embedding + bbox regression
• 111.5M params total, 48.4M trainable

Training

• Data: COCO 2017 (118K images, 80 classes)
• LR: 0.0001 (cosine decay)
• Batch size: 6
• Precision: bf16
• Best val_loss: 2.36 (epoch 2)
• Hardware: NVIDIA L4 23GB

Exported Formats

Format	File	Size
PyTorch	pytorch/loki_v1.pth	792MB
SafeTensors	pytorch/loki_v1.safetensors	426MB
Text embeddings	pytorch/text_embeddings.pt	83KB
Checkpoint (resume)	checkpoints/best.pth	792MB
Training config	configs/training.toml	—
Training history	logs/training_history.json	—

Usage

# Requires HDINO repo: https://github.com/HaoZ416/HDINO
from safetensors.torch import load_file
weights = load_file("pytorch/loki_v1.safetensors")
model.load_state_dict(weights, strict=False)

ANIMA Module

• Codename: LOKI
• Tier: 2 Perception
• Wave: 6
• Downstream: TYR (tracking), FENRIR (manipulation)

License

Apache 2.0 — Robot Flow Labs / AIFLOW LABS LIMITED