| --- |
| license: apache-2.0 |
| tags: |
| - robotics |
| - haptics |
| - spatial-understanding |
| - touch-sensing |
| - force-estimation |
| pipeline_tag: robotics |
| --- |
| |
| # Motoko Spatial 1B |
|
|
| Motoko Spatial 1B is a foundation model for 3D haptic spatial understanding in robotics. It takes raw sensor array input from distributed touch sensors across a robot surface and outputs spatial force maps, contact region predictions, and pressure distribution fields. |
|
|
| ## Model Details |
|
|
| - **Model type:** 3D haptic spatial foundation model |
| - **Parameters:** 1B |
| - **Architecture:** Hybrid CNN + Transformer |
| - **Input:** 3D coordinate arrays and sensor pressure grids |
| - **Output:** Force field maps, contact region masks, and pressure heatmaps |
| - **License:** Apache-2.0 |
|
|
| ## Intended Use |
|
|
| Motoko Spatial 1B is designed for robotics systems that need dense touch and contact understanding from distributed tactile sensors. |
|
|
| Primary use cases include: |
|
|
| - Dexterous multi-finger manipulation |
| - Full-body robot touch sensing |
| - Terrain and surface contact mapping |
| - Collision detection |
| - Safe human-robot contact |
|
|
| ## Inputs |
|
|
| The model expects structured haptic sensor input containing: |
|
|
| - 3D sensor coordinates |
| - Pressure grid values |
| - Optional force and torque channels |
| - Sensor timing or sampling metadata when available |
|
|
| Raw haptic arrays should be converted into model input tensors with `preprocessor/feature_extractor.py`. |
|
|
| ## Outputs |
|
|
| The model produces spatial predictions for downstream robotics control and perception: |
|
|
| - Spatial force field maps |
| - Contact region masks |
| - Pressure distribution heatmaps |
|
|
| ## Repository Files |
|
|
| | File | Description | |
| | --- | --- | |
| | `config.json` | Architecture definition, including layers, attention heads, hidden size, channel count, and spatial dimensions. | |
| | `configs/sensor_config.yaml` | Sensor array layout, sampling rate, axes, channel names, and physical units. | |
| | `preprocessor/preprocessor_config.json` | Signal normalization, channel statistics, windowing, and resampling configuration. | |
| | `model/model.safetensors` | Actual trained model weights. The current scaffold contains a placeholder until trained weights are added. | |
| | `model/model.safetensors.index.json` | Weight index used for loading sharded or indexed safetensors weights. | |
| | `preprocessor/feature_extractor.py` | Converts raw haptic arrays into normalized model input tensors. | |
| | `tokenizer_config.json` | Signal tokenizer metadata for quantized or discretized haptic tokens. | |
| | `tokenizer.json` | Minimal tokenizer vocabulary placeholder. | |
| | `configs/training_config.yaml` | Training hyperparameters and checkpoint cadence. | |
| | `examples/inference.py` | Basic inference preprocessing example. | |
| | `examples/spatial_map.py` | Spatial force map construction example. | |
|
|
| ## Limitations |
|
|
| This repository is currently a minimal Hugging Face model scaffold. The included `model/model.safetensors` file is a placeholder and should be replaced with trained weights before production use. |
|
|
| ## Citation |
|
|
| Citation information will be added when a technical report or paper is available. |
|
|
