Amazon FARs OmniRetarget AI Robot Nails Parkour Blind ($13764 HUMANOID)

Look at how it jumps up onto the table, jumps back down. This is with five rewards and four domain randomization terms. No history, no curriculum. This is the most incredible backflip off the wall. And here’s a video of behind the scenes how it works completely blind. It does it with an interaction mesh that remembers the spatial and contact relationships between the robot, objects, and the terrain just like this. And here are the real world results. This is Sim2Real Transfer Zero Shot from Omni Retarget holding this box. Next, watch it crawl. It gets better and better.

On the left side, it’s touching its knees. On the right side, it’s actually only touching its toes. And in this next example, you can see that it improves each time that it tries to jump onto the table. Here it almost falters. Next, pops up with its knee. Now it just hops directly on top. And lastly, it does a 180 turn. It’s amazing. And it jumps over this box. And Omni Retarget is ultra scalable. That’s because it can take one demo and make multiple data sets out of it just like this. And another big deal is that again, it goes from Sim2Real Hardware.

You see the simulation on the top real hardware on the bottom. Also, they can change the sizes, for instance, of the boxes in simulation, and then it does it directly in the real world. They can also change other things like the sizes of environmental obstacles here, the box, you can see it climbing in simulation, then direct to real world hardware, climbs right up onto the table. And even better is that Omni Retarget is a general framework, meaning that it works for multiple robots. Here you can see the T1 on the left and the H1 on the right.

And because of this, Video Mimic has way less artifacts. That’s things like toes poking through the ground, knees going through the ground, you can see with Video Mimic, hands going through boxes, etc. Omni Retarget performs much better without any of these artifacts. And here are the results of everything, especially the proprioceptive aspect of it where it’s only using the spatial map to place this chair, figure out where the table and everything is, jump up on top, and then orient itself to be able to execute this flip. And Omni Retarget is actually open sourcing over four hours of their retargeted trajectories, so make sure to check out their GitHub website.

It has these different demos with these simulations of different robots doing all kinds of tasks, so you can check it out there. So is this real, or is it CGI? Make sure to like and subscribe if you think it’s real, and if it’s CGI, then leave a comment below and explain to me why you think that. And NVIDIA also unveiled some of its own unitary demos at Coral 2025. Here it showcased Groot N 1.6, the foundation model, and it’s doing more than the previous N1.5 bimanual stationary manipulation. Here it can actually understand these commands like bring me the healthiest snack, and it has to go over and determine which one that is.

Here you can see it walks over to where the food is, and if we just fast forward here to where it finds the plate, it uses its vision language action model, figure out which is the healthiest. It determines it’s the apple, and it does this using its customizable reasoning vision language action model. It can turn these vague instructions into step-by-step plans, and here it walks back to this lady. We’ll fast forward here, and you can see it doesn’t just hand the snack over to her. It actually throws it, which is a special touch there.

And in AI agent news, Microsoft just integrated its autonomous AI agents for vibe working in its Microsoft 365 co-pilot for its core office applications, including Excel, Word, and PowerPoint. And there are two major features here, Agent Mode and Office Agent. Starting with Agent Mode, this is directly embedded into Excel and Word, and it uses reasoning models from OpenAI, for instance, to autonomously handle complex tasks, including data analysis, financial model building, and proactive error checking in spreadsheets. And the system follows a self-correcting plan, act, and check loop, and it reaches 57.2 percent accuracy in Excel tasks, which is notably behind the 71.3 percent performance of human testers.

But the second feature here is Office Agent. This functions as a separate multi-agent system within the co-pilot chat interface, and it’s powered by Anthropic. And this agent specializes in comprehensive document creation, and it generates Word files, PowerPoint presentations, and it can even perform its own web research and content organization. And these autonomous functions are presently available at Microsoft 365 co-pilot personal, as well as family subscribers, thanks to the Frontier program. But a wider unscheduled public rollout is also anticipated for later this year. Plus, Google just updated its AI coding agent called Jules with a command line interface and a public API.

And Jules’ tool, CLI, uses Gemini 2.5 Pro. And the API released this week connects Jules to systems like Slack for bug fixes or CICD pipelines, and other updates include a file selector for chats, a memory feature for preferences, and secure environment variable handling. And Jules runs in the cloud, cloning repositories, running tests, and creating pull requests with audio summaries. And it’s already handled thousands of tasks requiring human approval. Last but not least, let’s take a look at Juyuan, or AGI Bot. They are now selling their Ling-C X2 Youth Edition robot for 98,000 RMB, or roughly $13,764 US dollars.

And this robot features between 27 to 31 degrees of freedom, and it has three degrees of freedom in its neck, with three degrees of freedom in its waist, five to seven degrees of freedom in its arms, and two degrees of freedom in its neck. But it also features an RGB camera as well as RGBD, as well as touch sensors and 3D slam laser. As for weight, the robot weighs between 35 to 37 kilograms and runs for about an hour and a half on a single battery charge. As for the Ling-C X2 Explorer Edition and the flagship edition, the prices are yet unknown.

But check out this video. But make sure to like and subscribe for more AI news, and thanks for watching. Bye. [tr:trw].