On Going Direct, host Jordan speaks with Chris Paxton, a roboticist at Agility Robotics, and Michael Cho, co-founder of Beat Robot and RoboPapers co-host, about robot learning, data, evaluation, and ways to get more robots into the world.
Michael describes years of reading AI papers and starting a sidewalk robot project with his brother during the pandemic. They made a low-cost, remotely controlled robot game that gathered data for urban navigation research and led to competitions such as the Earth Rover Challenge. Chris traces his path through robotics research and says his enduring interest is long-horizon planning, including home robots that can find and deliver items.
Chris says the central challenge is getting robots to work reliably in the real world, where deployment can generate data. Michael identifies robotics data and real-world evaluation as major bottlenecks: unlike digital models, robots cannot be tested quickly against shared benchmarks, and different robots and homes make comparisons difficult. Both caution that a video can show a robot succeeding at one task without proving it can generalize.
Michael calls teleoperation valuable because it captures real robot actions and human intent, but robots break and the data is costly to scale. Chris explains reinforcement learning as a robot learning through its own attempts and outcomes. They see human video as useful for breadth and diversity, but limited by missing action labels, occlusion, differences between human and robot bodies, and the difficulty of matching human touch.
Chris describes sim-to-real training: practicing tasks in simulation, then transferring the learned policy to a physical robot. Michael says simulation can get robots most of the way, but handcrafted scenarios and rewards can leave real-world edge cases out; he also sees potential in world models, which need substantial video and teleoperation data. Both favor combining data sources and putting more robots into real settings. Michael suggests entertainment could help deploy robots at scale, while Chris notes that data gathered during playful uses may also aid research.
Michael says RoboPapers began as a way to ask researchers questions and understand their work beyond what papers reveal. He brings a non-researcher’s questions, while Chris contributes technical expertise; they aim to keep conversations detailed and discuss research limitations. Chris says interviews have shifted some of his skepticism about egocentric video, while reinforcement learning and long-horizon reasoning are among the areas he finds exciting.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Michael Cho, founder of FrodoBots and Ultimate Fighting Bots, talks with The Humanoid Hub about robot entertainment, embodied AI research, and the BitRobot Network.
