Paper Link: https://tactx-corl.github.io/TACTX.pdf
Tactile policies usually only work wit h the sensor they were trained on
TACTX aligns 3 sensors that measure touch in different ways:
- Daimon DM-Tac W: vision-based outputs images
- eFlesh: magnetic, outputs a small vector of field readings
- FlexiTac: resistive, outputs a pressure grid
Training
2 sensors were mounted on opposite fingers of a Franka gripper, 3rd sensor was swapped in for other runs which covered all 3 sensor pairs in both left/right orders (6 setups)
They collected 2670 trajectories, about 145k frames, 20% held out for validation
Made look-alike housings so sensors could not be distinguished by their casing
Loss:
- Reconstruction (L1): each sensor rebuilds its own reading, and also its partner’s reading from shared latent
- Contrastive alignment: Two sensors’ readings of the same contact are pulled together
- KL regularization: Each sensor’s latent distribution is kept close to a standard normal
- Warm-up: Contrastive and KL weights ramp up from zero, so the model learns to reconstruct
Settings: Adam, learning rate 1e-4, batch size 64, 300 epochs