Robótica Cognitiva

Welcome to the Seminar on Cognitive Robotics

Introduction of Speakers

• The seminar is hosted by CTIC, a research center at UCR, focusing on information and communication technologies.

• Dr. Federico Ruiz, an expert in cognitive robotics, is introduced as the main speaker. He holds degrees in electrical engineering from UCR and a PhD from the Technical University of Munich.

• Dr. Oscar Caravaca Mora is also acknowledged for facilitating this seminar within his robotics course.

Overview of Presentation

• Dr. Ruiz expresses gratitude for the invitation and outlines the agenda: discussing intelligence, cognition, artificial intelligence (AI), cognitive robotics, and specific laboratory projects related to these topics.

Understanding Intelligence and Cognition

Key Characteristics of Intelligence

• Discussion begins on what makes humans intelligent; characteristics include understanding others' actions and self-awareness.

• Learning ability is highlighted as crucial for performing new tasks independently or with assistance.

• Social interaction skills such as empathy and communication are identified as forms of intelligence.

Types of Intelligence

• Manual intelligence or manipulative skills are emphasized; they relate to physical activities that may have contributed to human communication abilities.

• Creativity is defined as combining existing parts into something novel or useful.

Manipulative Intelligence in Robotics

Laboratory Focus

• The lab focuses significantly on developing robots that assist people through manipulative tasks using robotic arms.

Video Demonstration

• A video showcases a person demonstrating high manual intelligence through balance and manipulation skills, illustrating concepts like friction and gravity.

The Process of Cognition

Cognitive Cycle Explained

• A block diagram illustrates cognition as a cycle involving perception, understanding the environment, learning from actions, reasoning, acting upon knowledge gained, and adapting based on outcomes.

Relationship Between Intelligence and Cognition

• Intelligence is described as a phenomenon while cognition refers to the processes involved in achieving intelligent behavior.

Historical Context of Artificial Intelligence

Early Developments

• Brief history highlights foundational discoveries in logic that led to AI development; Turing's contributions are noted regarding computational theory.

Turing Test

• The Turing Test concept is discussed—evaluating if machines can exhibit intelligent behavior indistinguishable from humans.

Neural Networks Evolution

Biological Inspiration

• Santiago Ramón y Cajal's discovery about neural networks laid groundwork for artificial neural networks leading to perceptrons developed by Minsky around 1969.

Challenges with Neural Networks

Historical Setbacks

• Neural networks faced rejection due to processing difficulties but saw resurgence with advancements in machine learning techniques during the 2000’s.

Symbolic vs Non-Symbolic AI

Distinction Made

Symbolic AI relies heavily on logical reasoning but struggles with real-world applications due to its lack of grounding in reality compared to non-symbolic approaches like machine learning which adapt better to imperfect environments.

Probabilistic Robotics Emergence

Addressing Uncertainty

• Probabilistic robotics emerged in response to challenges posed by uncertainty in sensor data allowing machines to operate effectively despite imperfections present in real-world scenarios.

Trends & Future Considerations

Current Landscape

• Observations indicate current trends resemble economic bubbles where initial excitement leads investors towards inflated expectations followed by potential downturn periods known as "AI winters."

Machine Learning Limitations

• While machine learning creates an illusion of intelligent machines capable of conversation (e.g., ChatGPT), it lacks true reasoning capabilities inherent only within classical symbolic AI frameworks.

Understanding Deep Learning and Its Limitations

The Nature of Learning in AI

• Deep learning and machine learning are primarily based on the concept of learning from observed data, which can lead to incongruences if the training data is not representative of reality.

• An example illustrates that a model trained predominantly on images of salmon fillets instead of whole fish resulted in incorrect outputs, highlighting the lack of reasoning capabilities in machines.

Creativity vs. Reasoning

• Despite its limitations, deep learning excels in creative tasks, such as generating logos or abstract art, showcasing its potential for creativity rather than logical reasoning.

• Caution is advised when making decisions based solely on machine-generated outputs due to their potential for unrealistic solutions.

Understanding Machine Intelligence

• Machines like ChatGPT do not possess true intelligence; they generate responses without understanding context or meaning, as noted by prominent researchers like Yann LeCun.

• Users should critically evaluate machine-generated content since these systems can "hallucinate" information—producing plausible but inaccurate results.

The Role of Data Quality in AI Outputs

Limitations Based on Training Data

• AI tools perform well with common knowledge due to abundant training data but struggle with advanced topics where data scarcity exists (e.g., graduate-level inquiries).

• As users delve into more specialized subjects, they may encounter inaccuracies because the models lack sufficient training data for those areas.

Expert Review Necessity

• There’s a contradiction between needing vast amounts of data for training and ensuring that this data is reliable and reviewed by experts.

AI's Application and Ethical Considerations

Creative Tasks vs. Critical Decision-Making

• While AI excels at creative tasks, it poses risks to jobs requiring critical thinking due to its ability to generate convincing yet potentially flawed content.

Legislative Responses to AI Challenges

• Discussions around regulating artificial intelligence have emerged within legislative bodies; however, caution is needed regarding reliance on machine-generated texts that may lack substance or references.

Cognitive Robotics: Understanding Movement and Intelligence

The Need for Sensory Input in Animals

• Animals require brains for movement and survival; unlike plants that remain stationary, animals need rapid processing capabilities provided by neural networks.

Evolutionary Perspective on Brain Development

• The evolution of animal brains correlates with their need for quick decision-making related to movement towards food sources or away from predators.

Early Robotics: A Historical Overview

Cheiky: One of the First Mobile Robots

• Cheiky was an early mobile robot developed in the 1970s using basic sensors and cameras to navigate environments while relying heavily on external processing power.

Localization Techniques

• Early robots faced challenges with localization due to sensor inaccuracies; innovative recalibration methods were developed using visual references similar to human navigation techniques.

Advancements in Robotic Mapping Techniques

SLAM Algorithm Development

• The introduction of SLAM (Simultaneous Localization and Mapping), refined through the late 20th century, allowed robots to create accurate maps while navigating uncertain environments using probabilistic models.

Autonomous Navigation and Mapping in Robotics

Introduction to SLAM (Simultaneous Localization and Mapping)

• The discussion begins with an example of a robot using SLAM for autonomous navigation, showcasing how it maps its environment effectively.

• Real-world applications of SLAM are highlighted, particularly in autonomous vehicles that navigate complex terrains like deserts, requiring 3D mapping capabilities.

Challenges in Autonomous Vehicle Technology

• The limitations of visual-based systems in vehicles are discussed, emphasizing the importance of distance-sensing technologies over simple camera systems.

• Tesla's reliance on visual techniques is critiqued due to their history of accidents caused by misidentifying objects, such as detecting painted bicycles as real ones.

Cost Implications of Advanced Sensors

• The high cost of advanced sensors used in reliable autonomous systems is noted as a barrier for widespread adoption among manufacturers like Tesla.

• A low-cost laser sensor example from a robotic vacuum cleaner illustrates the potential for creating affordable autonomous robots using open-source software.

Cognitive Capabilities and Human Interaction

• The importance of cognitive abilities in robots is emphasized, particularly their need to understand their environment through mapping and systematic navigation.

• An experiment monitoring human activities in kitchens demonstrates how machine learning can identify tasks based on object interactions.

Eye Tracking and Intent Recognition

• Eye-tracking technology is introduced as a means to infer human intentions by observing gaze patterns before actions are taken.

• Examples illustrate how eye movements can indicate future actions, providing insights into human behavior that could enhance robot interaction capabilities.

3D Perception and Semantic Understanding

• The concept of cognitive perception involves creating 3D maps that extract semantic information about the environment from point clouds generated by scans.

• Techniques for identifying planes within scanned data help robots understand spatial relationships between objects, enhancing their ability to navigate environments intelligently.

Object Recognition and Manipulation Strategies

• Robots can be programmed to recognize doors or furniture based on identified vertical planes, allowing them to interact with their surroundings more effectively.

• Machine learning techniques enable robots to learn typical movement patterns when navigating obstacles or performing tasks.

Advanced Robotic Control Mechanisms

• Soft control mechanisms allow robots to interact safely with humans by emulating spring-like behaviors during contact instead of applying rigid force.

• Demonstrations show how soft control enhances manipulation tasks while ensuring safety during interactions with people nearby.

This structured summary captures key discussions from the transcript regarding robotics' advancements in navigation, perception, interaction capabilities, and safety measures. Each bullet point links back to specific timestamps for further exploration.

Importance of Emulating Human Behavior in Robotics

The Need for Human-like Interaction

• Emulating human behavior is crucial for robots interacting with people, as it aligns their actions with human expectations.

• This emulation can enhance attention and safety, particularly in environments where robots operate alongside humans.

Cognitive Architecture in Robotics

• The laboratory focuses on developing robots for everyday tasks, utilizing cognitive architecture to guide their functions.

• Current projects include a smart kitchen and a small retail store, both designed to present unique challenges for robotic interaction.

Challenges of Operating in Uncontrolled Environments

Safety and Adaptability

• Robots must navigate complex environments filled with diverse objects while ensuring safety and smooth interactions.

• Unlike industrial robots that operate in controlled settings, these robots face unpredictable scenarios requiring advanced adaptability.

Robot Capabilities

• A robot's design includes high dexterity and the ability to manipulate various everyday objects safely.

• Equipped with advanced sensors (e.g., 3D cameras, thermal imaging), the robot can perceive its environment effectively.

Technical Specifications of the Robot

Hardware Features

• The robot features powerful computing capabilities with significant RAM and processing power to handle complex tasks.

• Demonstrations show the robot's ability to move its torso and neck while expressing emotions through facial features.

Software Integration

• Utilizes Robot Operating System (ROS), an open-source software platform that enhances robotic functionalities through various utilities.

Programming Robotic Actions Based on Cognition

Affordances and Mirror Neurons

• Research incorporates concepts like affordances—how objects can be used based on their properties—and mirror neurons that facilitate learning through observation.

Manipulation Techniques

• Initial work involves modeling object manipulation based on physical properties such as friction to enable autonomous control during tasks like pushing objects into desired positions.

Complexity of Object Manipulation

Challenges in Control Systems

• Manipulating objects introduces complexities due to non-linear dynamics; slight changes can lead to chaotic outcomes during movement.

Feedback Mechanisms

• The system employs feedback loops using predictive models to adjust actions based on real-time conditions encountered during manipulation tasks.

Understanding Affordances Through Robotics

Conceptual Framework

• Affordances inform how different actors interact with objects; this understanding is essential for programming effective robotic responses.

Predictive Modeling

• A predictive model allows the robot to anticipate necessary actions based on observed behaviors or intended manipulations of objects.

Designing Cognitive Systems for Robots

Task-Specific Models

• Each task requires tailored predictive models that activate upon detecting specific objects, enhancing operational efficiency through specialized libraries.

Future Developments

• Plans include expanding object recognition capabilities beyond basic shapes like boxes and cylinders towards more complex interactions involving liquids or spreading substances.

Human-Robot Interaction Design Considerations

Emotional Expression

• The design aims at creating relatable emotional expressions in robots, which positively influences children's engagement during lab visits.

Avoiding the Uncanny Valley Effect

• Striving for a balance between humanoid appearance without crossing into uncanny territory ensures better acceptance among users by avoiding lifelike but awkward designs.

Evolutionary Robotics: Optimizing Robot Design

Body Structure Optimization

• Investigating optimal body structures tailored specifically for task execution rather than mimicking human anatomy directly leads to improved functionality.

Simulation Testing

• Prior simulations help determine ideal configurations before physical construction, reducing costs associated with trial-and-error methods in robotics development.

Robotic Arm Simulation and Control

Overview of Robotic Testing

• The team conducted tests on a robotic arm by simulating various positions to assess asymmetry between different sides and orientations.

• Height adjustments were not tested due to the robot's torso capability, but simulations included this feature for practical use.

Results from Simulations

• The current results show that the robot operates effectively at a 45-degree angle, with limited sampling of four angles in each direction due to extensive computational requirements.

• A total of 30 million executions were performed over two weeks to determine optimal configurations for the robotic arm.

Evolutionary Robotics Approach

• The project is inspired by evolutionary robotics, focusing on assembling existing parts rather than creating a complete robot from scratch. This area remains underexplored in research.

• Current efforts involve collaboration with a master's student to explore assembly techniques further.

Soft Control Mechanisms in Robotics

Impedance Control Development

• The goal is to implement soft impedance control across the entire robot, currently limited to hands and arms, enhancing safety during operation.

• There is minimal research available on soft control mechanisms for mobile platforms carrying significant weight (400 kg). A lab colleague is developing torque sensors for this purpose.

Torque Sensor Implementation

• Torque sensors will be mounted on the robot's wheels, allowing it to interact safely with its environment through controlled movements akin to springs throughout its structure.

• Existing wheel designs lack safety compared to arms and hands; thus, new solutions are being developed specifically for heavy-load scenarios.

Custom Controller Development

Open Compliance Robot Controller

• Due to high costs associated with purchasing individual impedance-controlled joints (approximately $22,000 each), a custom controller named "Open Compliance Robot Controller" was created using extensometer technology for torque measurement.

• After several iterations, a functioning version has been integrated into humanoid robots within the lab setting.

Sensor Design and Calibration

• A mechanically designed load cell measures applied force through metal deflection using an extensometer; data capture systems are currently being calibrated and tested in real-time applications.

Mobile Platform Movement Demonstration

Omnidirectional Mobility

• The video showcases an omnidirectional platform utilizing mecanum wheels that allow lateral movement without turning—ideal for navigating tight spaces typical in human environments.

Q&A Session Insights

Audience Engagement

• Following the presentation, questions arose regarding potential graduation projects related to mechatronics and cognitive robotics; opportunities exist across various domains including firmware development and human interaction studies.

Collaboration Opportunities

• Participants expressed interest in collaborative projects within mechatronics or cognitive robotics fields; discussions about specific project directions were encouraged among attendees seeking involvement or contributions.

Virtual Reality Integration in Robotics

Exploring VR Applications

• Discussion highlighted how virtual reality can enhance robotic control experiences by simulating environments where users can interact with virtual objects while receiving haptic feedback from physical robots.

Research Potential

• Future research may focus on controlling robots remotely via augmented reality setups that simulate real-world interactions within virtual contexts.

Challenges of Simulation Accuracy

• While VR offers unique advantages for training and testing robotic responses, challenges remain regarding accurately simulating complex physical interactions like friction which often fall short in current simulators.

This structured summary captures key insights from the transcript while providing timestamps linked directly back to relevant sections of discussion for easy reference during study or review sessions.