Generative AI for Kids

Meet Twiggy, a stick insect (her nickname is just for our show), and RedMirror, a real six-legged robot from VISTEC in Thailand. Scientists from Tohoku University in Japan and VISTEC used an AI to watch only three or four steps of a stick insect walking. The AI didn't copy those steps. It worked out what the bug was trying to do, like a coach would. The robot then practiced chasing that goal: try, get a score, try again. It learned to walk in about an hour. In computer tests, it crossed bumpy ground it had never seen and kept walking on five legs after one leg was switched off. This week we also talk about why AI works best as a coach, not a copy machine, and our Kid Mission turns you into a Bug Coach. The wow picture: A skinny twig-like bug takes a few careful steps, and a robot about five times bigger learns from them. Later, one of its six legs is switched off, and it just keeps walking on the other five. The AI learned from an open dataset showing just 3–4 steps of the stick insect Medauroidea extradentata (2,000 frames covering 18 leg joints). The method is called adversarial inverse reinforcement learning. The AI works out the insect's goal (its "reward") instead of being told how to move each leg. Tohoku University says the robot learned to walk in just an hour, and that it learned three times faster than with a standard reward. When the learned goal was moved to a different robot body, that robot reached a useful walk in 70,000 training steps instead of 200,000 (Interesting Engineering, reporting on the paper). It was trained only on flat-ground data. On uneven terrain it switched to a more wave-like leg pattern, and its forward speed dropped only slightly. With one of six legs disabled (simulated damage), it reorganized its remaining legs and shared the load differently to stay stable. A preliminary test on the physical RedMirror robot showed walking that looked like the simulations. Quote from Associate Professor Dai Owaki (Tohoku University): "We never told the robot how to walk. We asked what the insect was trying to achieve, and let the robot chase the same thing entirely on its own." Owaki also said a few steps from one stick insect "were enough to find a principle that works on a machine five times its size." Possible future use: legged robots that help at disaster sites where wheeled robots can't go. Paper: "From Insect Behavior to Transferable Robot Locomotion: Inferring Embodied Locomotor Principles from Limited Data via Adversarial Inverse Reinforcement Learning," by Yuchen Wang, Chuthong Thirawat, Mitsuhiro Hayashibe, Poramate Manoonpong, and Dai Owaki, published in Bioinspiration & Biomimetics (IOP Publishing). DOI: 10.1088/1748-3190/ae901f Sources: Tohoku University Research News, "Six-Legged Robot Learns to Walk from a Stick Insect" (2026-08-27): https://www.tohoku.ac.jp/en/press/%20six_legged_robot_learns_to_walk_from_insect.html Tohoku University press release, Japanese version (2026-08-19): https://www.tohoku.ac.jp/japanese/2026/08/press20260819-04-locomotion.html Peer-reviewed paper, Bioinspiration & Biomimetics, IOP Publishing (2026): https://iopscience.iop.org/article/10.1088/1748-3190/ae901f Interesting Engineering, "New six-legged robot mimics stick insect to master walking across uneven terrain" (2026-08-28): https://interestingengineering.com/ai-robotics/new-six-legged-robot-mimics-stick-insect iHLS, "AI Gives This Six-Legged Robot an Insect-Inspired Way to Walk" (2026-09-23): https://i-hls.com/archives/138770 Kid Mission: Be a Bug Coach Get down on your hands and knees and crawl across the room. That's four legs! Now tuck one hand behind your back and try again. (Go slowly on a soft rug or carpet.) What did you change to keep going? Did you lean, scoot, or hop? Tell a grown-up your new plan. Bonus: if you can safely watch an ant or beetle outside (just watch, don't touch), count how many legs are touching the ground at one time. You'll be doing what RedMirror did: keep the goal, find a new way.

A humanoid robot walks into a living room it has never seen — toys on the floor, a sock on the couch, a stuffed dinosaur under the table — and gets to work with its whole body. This episode of Generative AI for Kids follows Helix 2.5, Figure's newest neural network (announced September 17, 2026). After pretraining on Index, Figure's huge dataset of people showing how they move and do chores, the robot tried three jobs in homes it had never practiced in. The wow picture: They took Helix 2.5 into 30 Bay Area homes with zero data collected in any of them. No fine-tuning. Unseen toys, towels, and beds. Tidy living rooms (toys into a basket). Fold towels into a basket. Make beds (pillows plus comforter). In a blind test with the same task data, success jumped from 9% without Index pretraining to 56% with it — over 6× better. Full-task only. No partial credit. The robot even self-corrects: steps back, repositions, walks around the bed to fix a fold. Coach, not copy machine. Not "robotics is solved" — first strong evidence that whole-body skills can transfer from human experience to new homes. Sources: Figure AI, Sept 17, 2026: https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization Kid Mission: Helix Home Challenge Pick five toys in a room that is not your usual tidy spot. Put them into a basket, one by one. Or fold one towel carefully, corners together. If a fold is messy, step back, walk around, and try again. You just did Helix thinking. Learn from watching. Then help in a new place.

A robot arm sits on a messy desk. Blocks, a water bottle, a backpack — and a bright toy past the clutter. Pretty close is not good enough: a path that almost misses still knocks things over. This episode of Generative AI for Kids follows HardFlow, a new method from MIT researchers (published September 14, 2026) that helps generative AI invent clever paths while obeying nonnegotiable safety rules. The wow picture: Older helpers often force every tiny sketch step to stay perfect, which can make paths shy and slow. HardFlow gives the model more freedom while it explores, then enforces hard constraints on the final answer — no collisions, plus better quality (like the shortest legal path). In tests on robotic arms, mazes, and image editing, it hit perfect constraint satisfaction and found better solutions than other methods. Humans write the rules. The AI searches. Coach, not copy machine. Sources: MIT News, Sept 14, 2026: https://news.mit.edu/2026/new-method-enables-ai-safety-critical-situations-0914 Kid Mission: Desk Obstacle Challenge Clear a table corner. Place three obstacles and one sticker prize. Draw a top view. Invent three paths for an imaginary robot finger. Cross out any that touch obstacles. Circle the shortest safe path. You just did HardFlow thinking. Draft audio for this week's show. Not published yet.

After a year stuck, an AI coach called Get Physics Done found a Sun-dip path to Alpha Centauri in about a week. Slow down, swing closer than Mercury, fire near the Sun so panels stay small. Kid Mission: Near-Star Map Challenge. Draw Earth, Sun, Alpha Centauri. Invent three paths. Pick one and say why it saves power or weight. https://www.technologyreview.com/2026/09/01/1143247/ai-interstellar-journey-alpha-centauri/

A tiny robot duck learned to walk, get back up, and roller-skate. Not from a homework sheet. From practice. Microduck, from Hugging Face and Pollen Robotics (Aug 27). Same family as Reachy Mini. Four crayon colors: Cream, Graphite, Lavender, Sky. How it learns: reinforcement learning. Try. Score. Try again. Like a bike. Thousands of pretend ducks practice in a video-game Earth, then the winning moves go onto the real desk. Kid Mission: Duck Coach Challenge. Write a robot recipe. Have someone follow it exactly. When they get stuck, change the recipe and try again. https://pollen-robotics.com/microduck/blog/introducing-microduck/