Deepak Trivedi | GE Aerospace: Why does scaling a robot's degrees of freedom from 10 to 100 increase control complexity by 10,000x?

Can we eliminate CPU processing latency by physically embodying neural networks inside tactile skins?

Integrating high-density tactile pressure sensors with memristor arrays allows us to perform in-material computing directly at the point of contact. By fabricating physical neural networks within flexible substrates, raw sensor data can be processed on-site, entirely eliminating the need to transmit signals to an external processor. This merges the sensor and the computer into a single physical entity.

This localized architecture has already demonstrated the ability to execute noise reduction, real-time edge detection, and image sharpening directly within the sensor skin. Because the physical structure itself performs the mathematics of the neural network, the system acts as its own analog computer. This represents a massive shift from traditional systems that rely on computer vision algorithms run on distant GPUs.

Moving processing to the material layer drops latency down to single-digit milliseconds. This rapid feedback loop is critical for contact-rich interactions, such as grasping objects with unknown compliance or friction. Without this localized processing speed, robots cannot react quickly enough to prevent slipping or damage during complex manipulation tasks.

In this short video, you can learn:
* The fabrication architecture of high-density tactile sensors integrated with memristor arrays for in-material computing.
* How physical neural networks execute signal processing tasks like edge detection and noise reduction locally without digital CPUs.
* The critical role of single-digit millisecond latency in managing real-time, contact-rich robot interactions.

📋 **Clip Abstract** This clip explores the cutting-edge integration of memristor networks within flexible tactile sensor arrays to achieve true in-material computing. It demonstrates how physical neural networks can perform complex image processing and signal sharpening locally, dramatically reducing latency to the single-digit millisecond level.
🔗 Link in comments 👇

How does reservoir computing allow soft robots to process massive sensory loads with 99% less data transmission?

Reservoir computing is emerging as a powerful framework to exploit the rich, non-linear physical dynamics of a soft robot's structure. Instead of mapping every point sensor directly to an ASIC, the physical body itself acts as the initial computational reservoir, absorbing and filtering environmental interactions. This allows the structural physics of the robot to do the heavy mathematical lifting.

This approach mimics the biological split between local reflexes and central perception. By performing the heavy lifting of signal processing within the local physical topology, up to 99% of raw sensory data can be compressed into high-level features right at the edge. The brain or central CPU only receives the highly filtered, high-value information it needs to make top-level decisions.

As a result, the communication bandwidth needed between the robot's extremities and its central controller is reduced by orders of magnitude. This hybrid architecture drastically cuts down on power consumption, system weight, and communication latency. It provides a realistic pathway for scaling soft robotics without hitting the limits of silicon-based computing.

In this short video, you can learn:
* The conceptual framework of reservoir computing and how it exploits structural dynamics for physical signal processing.
* How local material loops can filter and compress up to 99% of raw sensory data before it reaches a central CPU.
* The architectural split between low-level local reflexes and high-level centralized perception in advanced soft robotics.

📋 **Clip Abstract** This clip explains how reservoir computing leverages the physical properties of a soft robot's body to perform initial computational steps. By compressing raw sensory data locally, this method reduces required communication bandwidth to a fraction of traditional centralized architectures.
🔗 Link in comments 👇