Boston Dynamics just showed the shop floor where it builds Atlas. Machining, assembly, and repairs all happen under one roof. When suppliers can’t meet the required tolerances, they make critical parts in-house. Engineers can tweak a design, machinists can make the part, and the team can fit and test it the same day. Building an Atlas currently takes about a day. The target is one hour next year, then eventually minutes. That pace is tied to Hyundai’s bigger plan. Hyundai Motor Group is targeting a US robot factory with annual capacity for 30,000 units by 2028, alongside plans to deploy more than 25,000 Atlas robots across Hyundai and Kia plants worldwide. All Atlas output for 2026 is already allocated to Hyundai’s RMAC training center and Google DeepMind. External customers are expected to start receiving robots in 2027. RMAC sits at Hyundai’s Metaplant in Georgia, giving Atlas a real automotive factory to work in. Initial tasks include sorting parts, with factory operations targeted for 2028 and assembly work to follow. The factory deployments could also feed data back into DeepMind’s models. Better models could help Atlas take on harder tasks, generating more real-world data in turn. Boston Dynamics also has an existing customer base for Spot and Stretch, plus its dealer and integrator network. Hyundai has discussed the possibility of using its car dealership network to sell robots, too. The plan still has to clear some big hurdles: manufacturing cost, reliability, and whether factory data can translate into consistent performance. Getting build time down from a day to an hour is one piece. The harder test is whether Atlas can do reliable work on the line.

Affordable humanoid robots share hands with humans. Most humanoid hands sit behind a paywall so high that only labs and big companies get to play. Yeah Hand, by Italian engineer Vittorio Lumare, tries the opposite. Mostly 3D-printed. Five servos per hand. Fishing-line tendons, like yours, with the motors in the forearm, not the fingers. PTFE guides keep friction down; in tests it opened and closed 800 cycles without getting close to overheating. It can lift up to 8 kg and put about 5 kg of grasp force into a power grip. Silicone pads for smooth stuff, tiny thumbnails for fine picks. Two flavors: a robotic version with ROS and mounting flanges for common arms, and an experimental prosthetic for research. A flex-sensor glove lets you mirror your own hand. On Tindie right now it's still in beta--kits from about $217, assembled around $581.

A humanoid robot that can catch a moving cup Most robots freeze the second something starts moving. Cup on a belt, shifting grasp, anything not stuck in place --and they miss. Long-WAM from NVIDIA (with MIT, HKU, UCSD) went after that. On a Unitree G1, stacking a cup rolling past on a conveyor: 19 out of 20. π0.5 and Fast-WAM? Zero. Even at 7.5 cm/s it still hit 90%, where both baselines blanked every try. What's different under the hood? It keeps ~19 seconds of what the robot just saw, quickly guesses what happens next (no full video render), then moves. Remember, imagine, act — in that order. Here's the part that actually matters: longer memory only helps if the model first learned to predict video forward in time. Do that, and RoboCasa GR-1 climbs to 78.7% with 19.2s of context. Skip it, and the extra history barely moves the needle. It runs on the edge too--RTX 5090 down to Jetson AGX Thor--as fast as 107 ms per action chunk. Factories, warehouses, any place where stuff doesn't sit still. This is the kind of memory humanoids have been missing.