What if robots could flip, climb, and crawl... completely blind? No cameras, no sensors, just pure AI smarts!
In this week's AI News, we're diving into Amazon's revolutionary Omni Retarget from Frontier AI & Robotics that's turning robots into acrobatic wizards. Plus, Nvidia's mind-reading bots, Microsoft's autonomous office agents, Google's coding supertool, and a $13K humanoid that's already started selling to consumers.
Amazon FAR team unveiled OmniRetarget, an innovative data generation engine for training complex whole-body loco-manipulation in humanoids.
This engine is designed to bridge the embodiment gap by using an interaction mesh that explicitly preserves critical spatial and contact relationships between the robot, terrain, and objects. By doing this, OmniRetarget transforms human motion capture data into kinematically feasible trajectories, generating over 9 hours of high-qualitydata that allows proprioceptive Reinforcement Learning (RL) policies to be trained efficiently.
This high-quality data enabled a Unitree G1 humanoid to successfully execute a complex, long-horizon dynamic sequence: carrying a chair, using it as a step to climb onto a platform, leaping off, and finishing with a parkour-style roll.
This entire 30-second sequence is drivenby a proprioceptive-only policy (no vision or LIDAR) trained with just 5 reward terms and simple domain randomizationm demonstrating significant progress in agile, human-like robot movement and complex scene interaction.