And one of the G1’s standout features across all four models is its joint articulation, with 23 to 43 degrees of freedom depending on the model. The robot’s knees, shoulders, elbows, wrists and ankles are all highly flexible for human-like movements. But the differences between these four new G1 robots begin with their computation and mobility. To start, the entry-level standard model offers a baseline of advanced spatial computation and serves as a solid, adaptable humanoid robot, with its initial price tag of $16,000 having been raised to $39,900. Then, for four grand more at a total price of $43,900, the Plus model upgrades its processing power to the NVIDIA Jetson Oren, delivering 100 trillion operations per second, making it a better fit for developers working on artificial intelligence and machine learning applications.

Next, the Smart model is priced at $53,900 due to improvements to its mobility and dexterity, with its hip joints being upgraded from 1 to 3 degrees of freedom. While its arms feature enhanced articulation, increasing single-arm freedom from 5 to 7 degrees of movement, allowing the robot to perform more human-like actions more smoothly for tasks requiring greater precision and flexibility. And at the very top of the lineup is the Ultimate model, a $65,900 showcase of Unitree’s most advanced technology. Building on the smart’s mobility enhancements, the Ultimate is equipped with DEX 3 Force-controlled, three-fingered hands that are capable of performing extremely delicate manipulations.

And while an optional haptic version of the hands is available for those requiring touch sensitivity, the standard hands already provide a significant edge for dexterous robotic tasks. But AI isn’t just transforming robots, as a new startup called Genmo just made waves in the video generation world with the release of Mochi 1. An open-source video model boasting a total of 10 billion parameters. And because it’s the largest publicly available AI model for video generation to date, Mochi 1 establishes new benchmarks in two key areas, motion quality and accuracy in following text-based instructions. But what truly sets Mochi 1 apart from previous video generation AI models is its ability to generate videos at 30 frames per second, with these clips lasting up to 5.4 seconds each.

And while this might still seem short, the level of realism it achieves is shocking. In fact, Genmo claims that Mochi 1 can even simulate intricate physical effects, such as the movement of liquids, fur and hair, all with a degree of realism that rivals more specialized, proprietary models. However, the current iteration is optimized for photorealistic content, meaning it’s not as well suited for producing animated or cartoon-like visuals. And Genmo acknowledges that distortions may occur during extreme movements, highlighting the model’s focus on controlled realistic scenarios. But the current videos being generated by Mochi 1 are still limited to 480p resolution, with Genmo planning to release an HD version supporting its 720p resolution videos later this year.

But what’s more impressive about Mochi 1 is its technical backbone of a new architecture called Asymmetric Diffusion Transformer, or ASIMDIT. This design separates the processing of text and visual content, with the visual component commanding about four times the resources of the text processing part, and by allocating more parameters to video generation, Mochi 1 ensures higher quality visuals while maintaining efficient language understanding. And unlike many current diffusion-based models, Mochi 1 employs only a single language model to process prompts, with this approach streamlining the system without sacrificing performance. In fact, early benchmarks suggest Mochi 1 not only outperforms its competitors in following text prompts, but also delivers more realistic motion for complex physical effects.

And in another move that aligns with the growing open-source ethos in AI research, Genmo made both its model weights and code for Mochi 1 available under the Apache 2.0 license, meaning that developers and researchers can access these resources on platforms like Hugging Face and GitHub immediately, enabling experimentation and collaboration within the broader AI community. Plus, for people who are curious to see Mochi 1 in action, Genmo even offers a free playground on its website, allowing users to try out the model and view community-generated examples, with each example including the text prompts it used so as to give more insight into how the model interprets instructions and produces its outputs.

But another new AI model called The Matrix is also redefining video generation, but now with infinite-length 720p streams featuring real-time frame-level control. This allows The Matrix to seamlessly blend simulated and real-world environments, creating highly interactive and realistic video content. And unlike traditional simulators, which require extensive manual setup, The Matrix learns from diverse data sources, including AAA games like Forza Horizon 5 and Cyberpunk 2077, and even real-world footage too. Amazingly, this allows The Matrix to generalize more effectively, simulating scenarios such as a BMW X3 driving indoors, even without any explicit training whatsoever for such tasks.

Plus, by being built on a video diffusion transformer, DIT, The Matrix uses the innovative Shift Window Denoise process model to manage attention over long sequences, enabling smooth, continuous video generation. Additionally, its interactive module processes real-time user inputs, ensuring precise control at speeds of 8 to 16 frames per second. And with such strong generalization, frame-level precision, and open source availability, The Matrix may also pave the way for new possibilities in video gaming, virtual reality, autonomous vehicle simulation, and much more. And finally, Video Generation Startup Runway ML has also just launched its own groundbreaking feature called Expand Video, enabling users to alter video aspect ratios by generating new content around the original footage.

Users can even combine expansions to simulate cinematic camera movements, such as crash zooms or pullback shots, turning static clips into dynamic sequences, and by using text prompts, the new tool ensures the utmost visual consistency while expanding the frames. Initially available for Gen 3 Alpha Turbo users, this feature will roll out gradually, with Runway ML also offering tutorials through its Runway Academy in order to help creators harness the tool’s full potential. Thank you for watching!
[tr:trw].