Ep#112: SlipSense and Touch2Trace

In this talk

1. 0:36

   A hardware and learning feedback loop

   Tawu says the team sees tactile dexterity as a joint hardware and machine-learning challenge. Its approach cycles through building sensors, testing them on robots, and using results to guide the next hardware design.

2. 2:33

   SlipSense combines tactile pressure and vibration

   SlipSense detects contact and slip using a 32×32 piezoresistive array at 240 Hz and an accelerometer at 8 kHz. In testing, 76% of slip events were identified within 23 milliseconds; combining the two sensor streams also raised macro F1 to about 95. Training on a WSG gripper transferred without fine-tuning to the Solo hand, and augmentation helped with differences between robot platforms.

3. 34:02

   Touch2Trace learns cable handling without vision

   Touch2Trace uses fingertip tactile data and joint positions to teach a robot hand to feed a cable through repeated pinch-and-curl motions. Ten minutes of teleoperated demonstrations supported a policy that traced 20 centimeters, while a proprioception-only policy failed. The team trained on one USB cable and reported zero-shot tests on other cables, with outcomes affected by cable stiffness, material, and size.

4. 46:40

   Design choices and open challenges

   For cable tracing, the team found better results with higher tactile resolution and control frequency, and a 15-frame history at 60 Hz worked best in its experiments. A compact transformer policy with a Gaussian mixture output ran at 60 Hz; freezing its pretrained tactile encoder helped with the small demonstration dataset. The team noted limits when cables move beyond reach or differ substantially from the training cable, and discussed future work combining touch with vision.