Ravi Pandya - Influence-Aware Safety for Human-Robot Interaction | Talking Robotics

The podcast explores influence-aware safety in human-robot interaction, particularly focusing on how robots can understand and manage their influence on human behavior to ensure safety. Ravi Pandya, a PhD candidate at Carnegie Mellon University, discusses his work on the Influence-Aware Safe Control Problem, emphasizing the importance of considering human intentions and dynamics when designing robot policies. He presents projects on multimodal safe control, where robots navigate uncertainty in human actions, and model-based conditional behavior prediction, enabling robots to switch between courtesy and influence objectives in collaborative tasks. The discussion highlights the challenges of creating safe and efficient human-robot interactions, including managing uncertainty, predicting human behavior, and adapting to different interaction styles. Ultimately, the goal is to develop robots that can proactively influence people while maintaining safety and achieving task efficiency.

Outlines

Part 1: Introduction, Context

Part 2: Theoretical Framework, Safety Modeling

Part 3: Behavior Prediction, Influence Strategies

Part 4: SLIDE Framework, Technical Implementation

Part 5: Q&A, Practical Challenges

Sign in to continue reading, translating and more.

Continue

Ravi Pandya - Influence-Aware Safety for Human-Robot Interaction

Talking Robotics

Part 1: Introduction, Context

Introduction to Ravi Pandya and Influence-Aware Safety for Human-Robot Interaction

The Importance of Influence-Awareness in AI and Robotics for Human Safety

Part 2: Theoretical Framework, Safety Modeling

Formalizing Influence-Aware Safety: The Influence-Aware Safe Control Problem

The Necessity of Safety in Robotics: Dynamics, Uncertainty, and Value Functions

Multimodal Safe Control for Human-Robot Interaction: Addressing Uncertainty in Human Actions

Computing Least Conservative Controls: Balancing Safety and Task Completion

Validating the Gaussian Mixture Model Assumption with Pedestrian Prediction Models

Part 3: Behavior Prediction, Influence Strategies

Model-Based Conditional Behavior Prediction: Learning Influence-Aware Models

Separating Inference of Prior and Posterior Goals: Courtesy vs. Influence Objectives

Simulation and User Study Results: Evaluating Influence-Aware Models

Part 4: SLIDE Framework, Technical Implementation

SLIDE: Safely Leveraging Influence in Dynamic Environments

Solving for a Safe Policy: Reach-Avoid Dynamic Game and Adversarial Reinforcement Learning

Training and Evaluation: Comparing SLIDE to Robust Approaches

Analyzing the Conditional Behavior Prediction Model and its Impact on Efficiency

General Framework for Interactive and Safe Robot Policies and Future Applications

Part 5: Q&A, Practical Challenges

Q\&A: Addressing Safety Concerns in Out-of-Distribution Scenarios

Q\&A: Managing Tractability in High-Dimensional Belief Spaces

Q\&A: Modeling Implicit Safety Preferences in Human Models

Q\&A: Formally Delineating Influence in Robot Plans

Q\&A: User Preference for Robots Influencing Behavior

Ravi Pandya - Influence-Aware Safety for Human-Robot Interaction

Talking Robotics

Part 1: Introduction, Context

00:09Introduction to Ravi Pandya and Influence-Aware Safety for Human-Robot Interaction

Introduction to Ravi Pandya and Influence-Aware Safety for Human-Robot Interaction

01:14The Importance of Influence-Awareness in AI and Robotics for Human Safety

The Importance of Influence-Awareness in AI and Robotics for Human Safety

Part 2: Theoretical Framework, Safety Modeling

04:06Formalizing Influence-Aware Safety: The Influence-Aware Safe Control Problem

Formalizing Influence-Aware Safety: The Influence-Aware Safe Control Problem

06:23The Necessity of Safety in Robotics: Dynamics, Uncertainty, and Value Functions

The Necessity of Safety in Robotics: Dynamics, Uncertainty, and Value Functions

09:21Multimodal Safe Control for Human-Robot Interaction: Addressing Uncertainty in Human Actions

Multimodal Safe Control for Human-Robot Interaction: Addressing Uncertainty in Human Actions

12:12Computing Least Conservative Controls: Balancing Safety and Task Completion

Computing Least Conservative Controls: Balancing Safety and Task Completion

15:35Validating the Gaussian Mixture Model Assumption with Pedestrian Prediction Models

Validating the Gaussian Mixture Model Assumption with Pedestrian Prediction Models

Part 3: Behavior Prediction, Influence Strategies

16:24Model-Based Conditional Behavior Prediction: Learning Influence-Aware Models

Model-Based Conditional Behavior Prediction: Learning Influence-Aware Models

18:03Separating Inference of Prior and Posterior Goals: Courtesy vs. Influence Objectives

Separating Inference of Prior and Posterior Goals: Courtesy vs. Influence Objectives

20:13Simulation and User Study Results: Evaluating Influence-Aware Models

Simulation and User Study Results: Evaluating Influence-Aware Models

Part 4: SLIDE Framework, Technical Implementation

24:26SLIDE: Safely Leveraging Influence in Dynamic Environments

SLIDE: Safely Leveraging Influence in Dynamic Environments

26:26Solving for a Safe Policy: Reach-Avoid Dynamic Game and Adversarial Reinforcement Learning

Solving for a Safe Policy: Reach-Avoid Dynamic Game and Adversarial Reinforcement Learning

29:23Training and Evaluation: Comparing SLIDE to Robust Approaches

Training and Evaluation: Comparing SLIDE to Robust Approaches

31:08Analyzing the Conditional Behavior Prediction Model and its Impact on Efficiency

Analyzing the Conditional Behavior Prediction Model and its Impact on Efficiency

33:39General Framework for Interactive and Safe Robot Policies and Future Applications

General Framework for Interactive and Safe Robot Policies and Future Applications

Part 5: Q&A, Practical Challenges

34:41Q\&A: Addressing Safety Concerns in Out-of-Distribution Scenarios

Q\&A: Addressing Safety Concerns in Out-of-Distribution Scenarios

37:05Q\&A: Managing Tractability in High-Dimensional Belief Spaces

Q\&A: Managing Tractability in High-Dimensional Belief Spaces

39:44Q\&A: Modeling Implicit Safety Preferences in Human Models

Q\&A: Modeling Implicit Safety Preferences in Human Models

42:41Q\&A: Formally Delineating Influence in Robot Plans

Q\&A: Formally Delineating Influence in Robot Plans

44:54Q\&A: User Preference for Robots Influencing Behavior

Q\&A: User Preference for Robots Influencing Behavior