IROS 2025 Keynotes - Perception and Sensors: Roberto Calandra

In this talk

1. 0:05

   Touch as a foundation for dexterity

   Calandra argues that touch helps people infer properties such as an object's material and inertia during grasping. He points to experiments where people took four to five times longer to light a match after their sense of touch was anesthetized. He agrees that robots need touch for human-level dexterity, while arguing that tactile hardware has made substantial progress.

2. 3:10

   A digital revolution in tactile sensing

   Calandra compares tactile sensing with the spread of telephones and digital cameras, which enabled new ways to record and process audio and images. He expects affordable, reliable tactile sensors to support a new field he calls touch processing, including digitizing, transferring, and analyzing touch.

3. 5:44

   Sensors built for different needs

   Calandra describes vision-based tactile sensors, which use cameras to measure how contact deforms an illuminated elastomer. DIGIT was designed to be easy to manufacture and open source; he reports spatial resolution of about 20 to 50 microns and says thousands are in use. Later designs target other needs: Evetac captures dynamic touch at up to 1 kilohertz, Pinky was used in prostate cancer research, and Digit 360 combines more than 16 sensing modalities.

4. 14:25

   Building tools to process touch

   Calandra says touch processing is much less developed than computer vision and will require collaborative work to identify useful representations and inductive biases. His group released PyTouch as an early machine learning library for tactile data. He also describes work on representations that transfer across sensors and tasks, and on linking tactile measurements to language descriptions such as soft or rough.

5. 17:44

   Touch improves robot performance

   Calandra says touch provides direct information about contact position and interaction forces, which matter for robot safety and manipulation. In one grasping study, vision and touch reached 83% success on unseen objects, compared with 56% for vision alone; in match lighting, the results were 80% with vision and touch versus 20% with vision. He also reports better results from combining vision and touch for in-hand object rotation, and argues that reliable real-world robotics will need performance beyond the roughly 80% that may suffice in a lab study.