The End Of Knowledge - Vault 2 - ICLR (2014-2023)

graph LR classDef cho fill:#f9d4d4, font-weight:bold, font-size:14px; classDef softRobots fill:#d4f9d4, font-weight:bold, font-size:14px; classDef exoGlove fill:#d4d4f9, font-weight:bold, font-size:14px; classDef videoNet fill:#f9f9d4, font-weight:bold, font-size:14px; classDef exoAbs fill:#f9d4f9, font-weight:bold, font-size:14px; classDef collaboration fill:#d4f9f9, font-weight:bold, font-size:14px; classDef machineLearning fill:#f9d4d4, font-weight:bold, font-size:14px; A[Kyu Jin Cho
ICLR 2021] --> B[Prof. Cho: bioinspired soft robots. 1] A --> C[Embodied intelligence: outsourcing computation to
morphology, materials. 2] C --> D[Water-jumping robot: flea-inspired
torque reversal. 3] C --> E[Ladybird beetle: origami-based foldable wings. 4] A --> F[Human body: 600+ muscles,
4000 tendons, 900 ligaments. 5] F --> G[SCI patient: eat independently,
not robot-fed. 6] F --> H[2010: wearable robots rigid exoskeletons. 7] H --> I[ExoGlove: fabric-based soft robotic glove. 8] I --> J[ExoGlove Poly: polymer, passive thumb,
improved wearability, hygiene. 9] A --> K[VideoNet: vision-based intention detection. 10] K --> L[VideoNet: learned object existence,
affected by hand angle. 11] I --> M[Fingertip force feedback needed for
gentle grasping. 12] M --> N[Bending Time Gradient RSTM:
improves force estimation. 13] I --> O[Topology optimization: compliant
origami finger structures. 14] A --> P[Many in wheelchairs have
respiratory weakness. 15] P --> Q[ExoAbs belt: mimics abdominal muscles. 16] Q --> R[ExoAbs: improved speech volume,
clarity in stroke patient. 17] A --> S[Combining physical embodied intelligence,
machine learning: faster, better solutions. 18] A --> T[Human body complex soft robots
enable physical human-robot interaction. 19] A --> U[Collaboration needed: soft robotics,
machine learning communities. 20] U --> V[Collaboration: respect, communication,
perseverance through failures. 21] U --> W[Innovative datasets key for
machine learning in robotics. 22] A --> X[Bioinspired robots: observe nature,
hypothesize, build, model, iterate. 23] A --> Y[Balance passive soft robot dynamics,
active control. 24] Y --> Z[ExoGlove: passive adaptation reduced
machine learning grasp errors. 25] A --> AA[Machine learning interpretability important
for robotic reliability, failsafes. 26] A --> AB[Prof. Cho: machine learning can
accelerate novel robot design. 27] A --> AC[Soft robots needed for safe,
natural human-robot interactions. 28] A --> AD[Soft robotics needs machine learning
collaboration to tackle challenges. 29] A --> AE[Prof. Cho: inspire machine learning community
to advance bioinspired soft robotics. 30] class A,B,F,X,AC cho; class C,D,E,H,I,J,O,Y,Z softRobots; class I,J,M,N,O exoGlove; class K,L videoNet; class P,Q,R exoAbs; class G,S,T,U,V,W,AD collaboration; class AA,AB,AE machineLearning;

Resume:

1.-Professor Kyu jin Cho designs soft bioinspired robots by learning from nature and extracting key principles rather than directly copying natural mechanisms.

2.-Embodied intelligence involves outsourcing computation to body morphology and material properties when designing robots, rather than relying only on internal control structures.

3.-Prof. Cho's lab built a water-jumping robot inspired by the flea's unique 'torque reversal' jumping mechanism to enable jumping on water.

4.-A ladybird beetle inspired origami-based foldable wings were developed for a jumping and gliding robot that can rapidly self-deploy and lock.

5.-Human body has over 600 muscles, 4000 tendons & 900 ligaments - still a lot to learn from to improve robots and quality of life.

6.-Met SCI patient who wanted to eat independently rather than be fed by robot, showing importance of enabling human independence.

7.-In 2010, wearable robots were rigid exoskeletons, but the human body already has a frame (bones) and just needs muscle assistance.

8.-Developed ExoGlove, a fabric-based soft robotic glove with artificial tendons routed through small conduits to mimic the human tendon pulley system.

9.-ExoGlove Poly version made with polymer for water resistance, easy cleaning, reduced motors, and passive thumb for improved wearability and hygiene.

10.-VideoNet vision-based intention detection uses arm behavior and hand-object interaction to predict grasping intent without needing biosignals.

11.-Experiments showed VideoNet learned object existence rather than specific objects, and is affected by attack angle of the hand.

12.-Fingertip force feedback is needed for gentle grasping; sensors on actuator can estimate force from tendon tension like human Golgi tendon organs.

13.-Bending Time Gradient RSTM algorithm uses recent cable sheath bending angle data to account for nonlinearities and improve force estimation.

14.-Topology optimization, typically used for rigid structures, was applied to design compliant origami finger structures with better wearability.

15.-Many in wheelchairs have respiratory weakness; cough assist is needed to expel mucus and prevent potentially fatal mucus buildup.

16.-ExoAbs belt mimics abdominal muscles to provide customized compression for cough, breathing and speech assistance in a portable device.

17.-Pre-trial showed ExoAbs could significantly improve speech volume and clarity in a stroke patient by assisting exhalation force.

18.-Combining physical embodied intelligence in robots with machine learning could enable solving important real-world problems faster and better.

19.-The human body is very complex with many health issues; soft robots enabling physical human-robot interaction are important to study.

20.-More collaboration is needed between the soft robotics and machine learning communities to advance the field and improve people's lives.

21.-Collaborating across disciplines requires mutual respect, open communication, willingness to ask basic questions, and perseverance through failures.

22.-Identifying innovative datasets that capture new types of information is key when applying machine learning to novel robotics domains.

23.-Designing bioinspired robots involves observing nature, hypothesizing key principles, finding biological matches, building prototypes, modeling, and iterative experimentation and redesign.

24.-Balance between passive soft robot dynamics and active control depends on predicting user intent to determine needed compensation forces.

25.-Machine learning errors in grasp intent prediction were reduced by relying on passive adaptation from embodied intelligence in the ExoGlove.

26.-Interpretability of machine learning models is important for robotic applications to ensure reliability and failsafe operation, given data limitations.

27.-Prof. Cho hopes the machine learning community can help accelerate the design of novel robots by learning from biological principles.

28.-Soft robots are needed for safe and natural human-robot interactions in wearable devices, assistive technologies, manufacturing, and everyday life.

29.-The soft robotics community needs more collaboration with machine learning experts to tackle the challenging problems in the field.

30.-Prof. Cho aims to inspire the machine learning community to work on interesting problems in bioinspired soft robotics to advance the field.

Knowledge Vault built byDavid Vivancos 2024