Episode Summary
Executive Summary: Ayana Howard traces her robotics-and-AI career from early neural networks and human-demonstration learning to today’s work on pediatric robotics, embodied AI, and trust in autonomous systems. The conversation centers on how people overtrust robots, why timing and transparency matter when conveying uncertainty, and how trust is intentionally built in child therapy applications to improve outcomes.
Main Topics: Early AI and robotics origins (Priority: 5/5): Howard describes herself as an "old school" AI practitioner who began coding neural networks in the early 1990s, using AI mainly to make robots more intelligent rather than as a standalone research identity. Human-demonstration learning for deformable object manipulation (Priority: 5/5): Her PhD work at NASA JPL focused on teaching robot manipulators to grasp deformable objects like pillows and sheets by learning from human demonstrations and limited sensor data, then using a neural network to map object shape to force. Pediatric robotics for therapy and coaching (Priority: 5/5): Howard’s lab develops small humanoid robots that act as coaches/therapists for children with special needs, especially children with cerebral palsy, using emotion, gaze, facial expression, and motion tracking to drive engagement and adherence. Overtrust in robots and embodied agents (Priority: 5/5): A major research thread is that people often follow robot guidance even when robots visibly make mistakes, suggesting humans place robots in an expert-like state and overtrust them. Uncertainty communication and transparency (Priority: 4/5): Howard argues robots should communicate uncertainty in ways that are timely and actionable, not merely numerically transparent; the goal is to prompt human attention at the right moment without overwhelming the user. Autonomous vehicles and real-world risk (Priority: 4/5): She extends the overtrust problem to self-driving cars and other autonomous road systems, arguing that researchers should mitigate trust failures before major accidents damage public confidence. Embodied AI vs. virtual AI (Priority: 3/5): Howard sees strong overlap between embodied AI and other AI agents, but emphasizes that the physical world introduces unique constraints, errors, and adaptation challenges that pure virtual agents do not face.
Key Arguments: Her early AI work was not about hype or labels; it was a practical method for making robots smarter through learned models and human expertise. Limited-data neural networks were already useful in the 1990s, even with only about 10 objects, because the problem was structure and learning from demonstration, not scale. People can overtrust robots even when the robot is clearly malfunctioning, which contradicts the intuitive idea that obvious robot failure should break trust. Trust in robots seems less about generic authority and more about assigning robots an expert role in the situation. The same human behavior can be interpreted differently when delivered by a robot versus a person; the perception of the agent matters even when task performance is nearly identical. Uncertainty should be communicated contextually and at the right time; raw percentages like "80%" are often less effective than concise actionable warnings. Pediatric robotics requires intentional trust and bonding because the goal is improved long-term therapeutic outcomes, not just task completion. For vulnerable populations like children with special needs, introducing mistakes to study trust can be ethically inappropriate because it may affect behavior and therapy outcomes. Embodied AI and robotics now overlap heavily because compute and AI capabilities are strong enough to run advanced algorithms on physical systems. Future work should focus on measurable trust factors, demographic/behavioral susceptibility, and mitigation strategies before autonomous systems become too embedded in daily life.
Data Points: Year of first neural network coding: 1994 - Howard says she coded her first neural network in 1994. Number of deformable objects in thesis experiments: 10 objects - Her thesis work on grasping deformable objects used only about 10 objects, which she describes as a lot at the time. Therapy study duration: 8 weeks - Her pediatric robotics studies reportedly ran for eight weeks, which she notes is not truly long term. Robot size: roughly 18 inches to 2 feet - She describes the humanoid therapy robots as small, around 18 inches to two feet tall. Age range example for trust susceptibility: 16 to 20 - She speculates future models may identify demographic groups such as teenagers aged 16 to 20 as more susceptible to trust effects. Human confidence in robot guidance: People followed robot guidance despite visible mistakes - In evacuation simulations and real-world tests, participants often stayed with or followed broken or failing robots.
Pivotal Quotes: "I've been doing this since 1990, oh, 1994. I think I coded up my first neural network." — Ayana Howard: Howard explains her long history in AI and frames herself as an early practitioner. "We have evidence that people overtrust robots." — Ayana Howard: She summarizes the central finding from her trust research on evacuation and guidance scenarios. "It's not about being fully transparent, it's about providing the information that we need at the time that it's needed." — Ayana Howard: Howard describes her view on how AI systems should communicate uncertainty and decision-relevant information.
Implications: Robotics systems need human-centered trust design, not just technical accuracy. Future products should convey uncertainty contextually, and high-stakes domains like therapy and driving must address overtrust before deployment scales.