SharpaWave Deform Encoder
DeformEncoder converts a preprocessed single-channel scalar deformation image
into a compact learned tactile feature with shape [B, 512, 1, 1]. The feature
can be flattened to [B, 512] for downstream tasks.
The unified checkpoint also includes DeformDecoder parameters. The decoder
and DeformAutoencoder are provided only to demonstrate deformation
reconstruction from the compact feature.
Tensor Shapes
| Operation | Input | Output |
|---|---|---|
| Encoder | [B, 1, 240, 240] |
[B, 512, 1, 1] |
| Flatten feature | [B, 512, 1, 1] |
[B, 512] |
| Autoencoder | [B, 1, 240, 240] |
[B, 1, 240, 240] |
The input is a preprocessed scalar deformation image, not a raw RGB camera image.
Usage
Download the checkpoint and use load_encoder() from the source repository:
import torch
from huggingface_hub import hf_hub_download
from sharpawave_deform_encoder import load_encoder
checkpoint = hf_hub_download(
repo_id="Sharpa-Robotics/sharpawave-deform-encoder",
filename="sharpawave_deform_autoencoder.safetensors",
)
encoder = load_encoder(checkpoint, "cpu")
deform = torch.zeros(1, 1, 240, 240)
with torch.inference_mode():
feature = encoder(deform) # [1, 512, 1, 1]
Source code: https://github.com/sharpa-robotics/sharpawave-deform-encoder
Checkpoint
The single SafeTensors file contains both encoder and reconstruction-only
decoder parameters. load_encoder() reads only the encoder tensors;
load_autoencoder() loads the complete demonstration model.
The published checkpoint was trained from random initialization without upstream pretrained weights.
The SHA-256 digest is recorded in SHA256SUMS.
Limitations
The encoder expects the documented 240-by-240 scalar input representation.
License
Developed by Sharpa Group. Licensed under Apache License 2.0. See LICENSE and
THIRD_PARTY_NOTICES.md.