𝗣𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗔𝗜 𝗗𝗮𝗶𝗹𝘆: 𝗧𝗵𝗲 𝗥𝗮𝗰𝗲 𝗳𝗼𝗿 𝗛𝘂𝗺𝗮𝗻𝗼𝗶𝗱 𝗠𝗮𝘀𝘀 𝗣𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻
The physical AI industry is moving from labs to factories. Here are the key updates you need to know.
𝗠𝗮𝗷𝗼𝗿 𝗠𝗼𝘃𝗲𝗺𝗲𝗻𝘁𝘀
• Hyundai is buying the rest of Boston Dynamics for $325 million. This gives them full control. They plan to mass produce the Atlas robot by 2026.
• Galaxy Robotics released a new model called AstraBrain-WBC 0.5. They trained it on 2 billion frames of human motion. They claim this shows scaling laws in robot control, similar to how GPT works for text.
• Momenta is moving its IPO to Hong Kong. They want to raise $1 billion for autonomous driving tech.
• General Intuition is in talks to raise $300 million. Their valuation may exceed $2 billion. They use gaming data to train world models.
𝗧𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝗕𝗿𝗲𝗮𝗸𝘁𝗵𝗿𝗼𝘂𝗴𝗵𝘀
• Physical Atari: Researchers are moving reinforcement learning from simulations to real robots. They built a robot to play Atari games using real controllers and cameras.
• World Models: A new paper suggests current world models lack a persistent state. They forget things once the camera looks away.
• Image-Based Action: New research shows that using image editing models can be faster and cheaper than generating full videos for robot actions.
• Human Data: Using videos of humans performing tasks can be better and cheaper than using expensive robot teleoperation data.
• Dexterous Hands: AGILINK has reached unicorn status. They are already shipping thousands of robotic hands every quarter.
𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆 𝗧𝗿𝗲𝗻𝗱𝘀
• Automation vs Jobs: GM added 50 robots to a factory after laying off 1,000 workers.
• Logistics: Volvo self-driving trucks are now hauling real freight in Texas.
• Hardware: While China leads in humanoid robot shipments, most expensive parts like sensors and reducers still come from overseas.
Source: https://dev.to/future_x/futurex-physical-ai-daily-issue-33-0620-37db
Optional learning community: https://t.me/GyaanSetuAi