Ep#88: DexSkin: High-Coverage Conformable Robotic Skin for Learning Contact-Rich Manipulation
Suzannah Wistreich, incoming Stanford PhD student, and Baiyu Shi, a third-year Stanford mechanical engineering PhD student, discuss DexSkin, their conformable tactile sensor, and its manipulation experiments with RoboPapers hosts and audience members.
In this talk
- 1:01
Why DexSkin uses conformable coverage
Wistreich describes a flexible tactile sensor designed to cover rounded fingertips, arms, or other curved surfaces. Shi explains that its soft SEBS material and kirigami-style structure can stretch and fold into a dome, where sheet-based sensors may need molding. The team aims to combine broad coverage with spatial and force sensing.
- 6:18
Sensor design and practical limits
Shi says structured materials help the sensor recover quickly, with signals fading in well under 100 milliseconds after force is removed. The team emphasizes interpretable, individually addressable contacts; reported tactile resolution is about three to four millimeters. The sensor is less suited to tiny parts than optical sensors, and the open-source version is not watertight for submersion.
- 15:08
Building and calibrating the hardware
Shi estimates material costs at under $10 per hand and the circuit board at roughly $30 to $40. A stretchable hand version takes about an hour to cut and assemble, while casting a dielectric for the flexible version can take one to two hours. After testing 10,000 cycles, the team reports up to 7.5% to 8% change; a three-minute calibration helped recover about 90% of policy performance when transferring to a newer sensor.
- 25:04
Learning contact-rich manipulation
The team trained Diffusion Policies from 50 real-world demonstrations per task. In a pen reorientation task, full tactile signals helped the robot detect and correct perturbations that policies without tactile sensing or with spatially averaged signals missed. In a rubber-band packaging task, tactile feedback helped distinguish damaged bands and supported both selection and packaging.
- 38:43
Residual learning and open-source plans
For fragile blueberry handling, the team added a tactile-based residual policy to an imitation policy, using force thresholds and penalties for deviating from the original actions. Across just under 200 training episodes, the policy learned gentler handling than a random residual policy. The team says it has released hardware, firmware, guides, and data, and is preparing additional sensor formats and designs for larger taxel counts.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.