Motoko Embedding 1B is a foundation embedding model for haptic signal representation in robotics. It encodes raw force, torque, pressure, and vibration signals into rich fixed-dimension vector embeddings for retrieval, search, and cross-modal fusion.

Model Summary

Model type: Encoder-only Transformer
Parameters: 1B
Input: Force, torque, pressure, vibration sequences
Output: Fixed-dimension embedding vectors
License: Apache 2.0

Intended Uses

Semantic search over haptic datasets
Cross-modal alignment with vision and language
Haptic RAG pipelines for robotic agents
Dataset indexing and similarity clustering
Downstream fine-tuning with LoRA adapters

Architecture

Motoko Embedding 1B uses a signal-aware preprocessing stack followed by an encoder-only Transformer. Multichannel sensor streams are windowed, normalized, projected into token embeddings, and aggregated into a single fixed-size embedding representation.

Key design points:

Temporal patching over multiaxis haptic sequences
Rotary position embeddings for long-context signal modeling
Mean pooling over the final hidden states for embedding extraction
Optional projection head for cross-modal alignment

Input Format

The model expects synchronized haptic sequences containing one or more of the following modalities:

Force
Torque
Pressure
Vibration

Default sensor assumptions are defined in configs/sensor_config.yaml. Signal normalization and windowing parameters are defined in preprocessor/preprocessor_config.json.

Repository Layout

text

.
├── README.md
├── config.json
├── tokenizer_config.json
├── tokenizer.json
├── model/
│   ├── model.safetensors
│   └── model.safetensors.index.json
├── preprocessor/
│   ├── preprocessor_config.json
│   └── feature_extractor.py
├── configs/
│   ├── training_config.yaml
│   └── sensor_config.yaml
├── examples/
│   ├── inference.py
│   ├── embedding_search.py
│   └── cross_modal.py
└── .gitattributes

Key Files

File	Purpose
`config.json`	Encoder architecture: layers, heads, hidden size, projection dimensions
`configs/sensor_config.yaml`	Sensor input specs: axes, sequence length, sampling rate
`preprocessor/preprocessor_config.json`	Signal normalization, windowing, padding behavior
`preprocessor/feature_extractor.py`	Converts raw haptic arrays into encoder-ready tensors
`examples/embedding_search.py`	Vector similarity search over haptic embeddings
`examples/cross_modal.py`	Aligns haptic embeddings with vision or language vectors

Usage

Load the processor

python

from preprocessor.feature_extractor import HapticFeatureExtractor

extractor = HapticFeatureExtractor.from_pretrained(".")

Basic embedding inference

python

import numpy as np

from preprocessor.feature_extractor import HapticFeatureExtractor

extractor = HapticFeatureExtractor.from_pretrained(".")
sample = np.random.randn(1024, 12).astype("float32")
features = extractor(sample)

print(features["input_values"].shape)
print(features["attention_mask"].shape)

See examples/inference.py for a complete example.

Training

Baseline training parameters are provided in configs/training_config.yaml. These values are intended as a starting point for pretraining or continued domain adaptation, not as a claim of the exact recipe used for a released checkpoint.

Limitations

Performance depends heavily on sensor calibration and synchronization quality.
Out-of-distribution hardware setups may require updated preprocessing statistics.
Cross-modal alignment quality depends on the paired supervision used during training.
This repository scaffold does not include production weights.

Citation

bibtex

@misc{motoko_embedding_1b,
  title        = {Motoko Embedding 1B},
  author       = {Motoko},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/}}
}

monodox/motoko-embedding-1-1b

Motoko Embedding 1B