Stanford Seminar - Towards trusted human-centric robot autonomy

()

Safe Control

Challenges in building trustworthy autonomous systems include safe control, addressing the question of how close is too close when humans and robots interact.
Different stakeholders make different assumptions about human behavior, leading to different definitions of safety.
The speaker introduces the concept of a "safety concept" that provides a measure of safety and safe/unsafe controls based on the world state.
Various safe control strategies are mentioned, including velocity obstacle, contingency planning, Hamilton-Jacobi reachability, safety force field, and responsibility-sensitive safety.
The speaker introduces Hamilton-Jacobi reachability, a mathematical framework that captures closed-loop interactions between agents.
The solution to the Hamilton-Jacobi-Isaacs equation, denoted as V, can be interpreted as a measure of safety, with the goal of maximizing safety and avoiding collisions.
The challenge lies in making reasonable assumptions about how other agents behave, as overly conservative assumptions can lead to impractical safety sets.
The speaker proposes learning a state-dependent control set from data, leveraging the assumption that humans tend to take safe controls.
Control barrier functions are introduced as a way to represent control invariance sets and parameterize the state-dependent control set.
The learning problem becomes a control barrier function learning problem, which can be solved efficiently using gradient descent.
By incorporating the learned control set into the Hamilton-Jacobi-Isaacs equation, novel safety concepts can be synthesized that are more practical and data-driven.

Interaction-Aware Planning

The speaker then presents work on interaction-aware planning to make interactions between humans and robots natural and seamless without explicit communication.
They argue that humans are self-preserving and will engage in joint collision avoidance, and propose a method for achieving safe and fluent human-robot interactions through simple changes to a standard trajectory optimization problem.
The speaker introduces the concepts of legibility and proactivity in robot motion and discusses how they can be encoded into a robot planning algorithm.
They also consider the fact that humans are not completely selfless and introduce the idea of inconvenience to limit the amount of inconvenience that an individual experiences during an interaction.
The research introduces the concept of a "markup factor" to encourage robots to take proactive and legible actions when avoiding collisions with humans.
A higher markup factor leads to earlier and faster rotation of the robot, making its intentions more apparent to humans.
The research also introduces an "inconvenience budget" to limit the amount of inconvenience caused to both the robot and humans during collision avoidance.
The proposed framework outperforms baseline methods in terms of fluency, fairness, and pro-social interactions.
The research demonstrates the effectiveness of the proposed framework through simulations and human-in-the-loop experiments.
The framework can be easily extended to incorporate multiple agents and obstacles, making it suitable for indoor and outdoor navigation.

Ongoing Challenges

The speaker emphasizes the importance of intentional data collection, focusing on safety-critical scenarios and human responses in those situations.
They propose using control theory to provide inductive bias for learning methods, rather than applying learning methods for control theory.
Addressing the curse of dimensionality in reachability methods is crucial for computational methods.
Testing algorithms with humans in the loop is expensive and challenging, especially in terms of developing performance metrics.