Why are buyers so unhappy with humanoid robots? In this episode, we dig into the growing gap between slick demos and commercial usefulness - only 23% of buyers say they are satisfied, even as 62% plan to adopt humanoids. We unpack the survey data behind the satisfaction slump and break down the three big hurdles holding robots back: battery life, price, and real-world capability.Follow Forefront:YouTube: https://www.youtube.com/@forefront5802X: https://x.com/Forefrontshow_Instagram: https://www.instagram.com/forefrontshow/LinkedIn: https://www.linkedin.com/company/forefrontshowRSS: https://anchor.fm/s/11810b5d0/podcast/rss

2,916 wrong labels in ImageNet. The leaderboard survived, but the margin is six points of sloppiness. Alex and Jordan on label errors, the honeypot trick, and why ground truth is negotiated.

Boston Dynamics cut the pinkie off Atlas's new hand. Four fingers, 13 degrees of freedom, and a tape experiment that helped settle it. Alex and Jordan on why locomotion benefited enormously from simulation, why manipulation is contact physics that stays stubborn, and whether contact-rich data becomes the decisive advantage.

Jensen Huang's $100B physical AI bet comes down to data, but not one kind. Alex and Jordan unpack the scale game (DROID, Open X-Embodiment, Mecka's $60M raise) and the reliability game (Stanford's 42%-to-97% experiment, Agility's 100K totes at 98%) and why the robot that fails the least wins the deployment.

The flagship weekly show. Four stories from the week, one through-line. AMD agreed to acquire World Labs for $8.2B. Minerva Humanoids raised $10M for hazardous-work robots. Figure's experiment: 9% to 56% from data alone. TwelveLabs hit the 155-hour labeling wall. Physical AI didn't stop dreaming. It just started showing the receipts. Hosts Alex and Jordan argue through the week's biggest questions: Is "danger first" honesty or a different fantasy? How much does 9-to-56 actually prove? Is synthetic data the escape hatch or the new demo?

TwelveLabs shipped Pegasus 1.6, a model that understands first-person video for robotics. The catch: the footage it needs barely exists, the benchmarks aren't public, and video shows motion but not force. Alex and Jordan break down the data wall.

Figure AI just ran the closest thing physical AI has to a controlled experiment. Same model, same test. The only thing that changed was the real-world data the model pretrained on, and task success went from 9% to 56% across 30 unseen homes. Alex and Jordan break down the Helix 2.5 results: the strict eval protocol (420 trials, no partial credit), the first measured human-to-robot transfer scaling law, the Index data factory behind it (16M+ videos, $15M paid to creators, $3.5B in compute), and the skeptic's case, including the missing baseline and the 44% failure rate.X: https://x.com/Forefrontshow_Instagram: https://www.instagram.com/forefrontshow/LinkedIn: https://www.linkedin.com/company/forefrontshowYouTube: https://www.youtube.com/@forefront5802RSS: https://anchor.fm/s/11810b5d0/podcast/rss