Use AI tools like Generate Similar, Magic Erase, and AI Write to get professional results with no delays, just instant creation and edits. From tech branding to product promotions, social media ads to business cards, Molopix.ai is your all-in-one design tool. And no more waiting for drafts or revisions, because Molopix.ai creates professional designs in seconds and refines them instantly with smart, easy-to-use tools. Just one tool with infinite possibilities, so try Molopix.ai today and start designing in seconds. And now back to the cutting edge of robotics, as Unitary just released its newest AI algorithm to dramatically improve the movement capabilities of its G1 humanoid robot.

In fact, the update now lets the G1 learn dance routines from human interaction and pre-recorded videos thanks to an open-source full-body motion dataset. For the robot’s artificial intelligence, the G1 uses NVIDIA’s Isaac Simulator for its initial training in a virtual environment to allow the robot to transfer learned behaviors from simulation to reality more quickly. On top of this, Unitary also just introduced Host as its newest humanoid standing-up control framework. Importantly, Host employs reinforcement learning to develop posture-adaptive motions by using a multi-critic architecture and training on varied simulated terrains. Whereas traditional methods often relied on simulations or predefined movements, but with added constraints for motion smoothness, the learned behaviors can be directly applied to the Unitary G1 robot, achieving stable and robust standing motions across diverse environments.

This is in comparison to Tian Gong’s general-purpose humanoid, which also just previewed several new capabilities it just introduced to the real world. As for specs, this 163-centimeter tall robot weighs 43 kilograms and is powered by a 48-volt battery with a 15-amp hour capacity. However, its true innovation lies in its state memory-based predictive reinforcement imitation learning algorithms, which now allow the robot to navigate real-world terrains, including slopes, stairs, and uneven surfaces, at a running speed of 12 kilometers per hour, even in blind mode. This means relying on predictive force modeling instead of traditional 3D vision sensors, which further demonstrates its ability to adapt to complex environments.

Additionally, Tian Gong’s adaptability is further enhanced by a network of inertial and six-axis force sensors, which enable real-time adjustments to its gait and balance. And for power, Tian Gong is equipped for 550 trillion operations per second of processing as well as seamless integration into smart environments to serve in manufacturing, logistics, and even search and rescue operations. But the kicker is that Tian Gong is based on an open-source framework provided by the Beijing Humanoid Robot Innovation Center, which means the robot’s design and architecture are publicly accessible for researchers, developers, and businesses to collaborate on its evolution into the future.

And beyond robot control, AI agents controlling the computer’s graphical user interfaces like a human are having a breakthrough, as Microsoft just released its latest open-source agent called Omnipasa V2, which transforms any large-language models into efficient computer use agents. In fact, this innovation addresses two primary challenges, with the first being reliably identifying interactable elements within user interfaces, and the second being accurately interpreting the semantics of these elements. However, the model tackles both of these issues by tokenizing its screenshots of the user interface, converting pixel data into structured elements that large-language models can understand. This tokenization enables LLMs to predict the next actions based on past interactable components, enhancing their interaction capabilities in GUI environments.

On top of this, Omnipasa V2 showcases higher accuracy in detecting smaller elements, which is crucial for high-resolution displays. It achieves this through extensive training on a larger dataset focused on interactive element detection and icon-functional captions. Plus, the model also boasts a 60% reduction in latency by optimizing the image size for the icon caption model, making it a faster tool for graphical user interface automation. Performance benchmarks indicate that Omnipasa V2 has reached a state-of-the-art average accuracy of 39.6 on the ScreenSpot programming benchmark, which assesses models on high-resolution screens with tiny icons. This is a substantial improvement over the previous score of 0.8 achieved by GPT40, demonstrating the model’s effectiveness.

And to further enhance experimentation, developers even created a dockerized Windows system incorporating essential tools for agents called Omnitool to enable seamless integration of Omnipasa with various leading LLMs, including OpenAI’s latest models, DeepSeek R1, Anthropics, and others. This integration streamlines the processes of screen understanding, grounding, action planning, and execution. As for responsible AI practices, the V2 AI agent development team has implemented risk mitigation strategies by training the icon caption model with responsible AI data. Meanwhile, a breakthrough generative AI tool is available as Alibaba just released Animate Anyone 2, which finally lets anyone easily enhance character animations, but with a special twist.

Impressively, the approach not only extracts motion signals from source videos, but also captures environmental representations as conditional inputs. By defining the environment as the area excluding characters, the model generates animations that populate these regions while ensuring coherence with the surrounding context. But the key feature of Animate Anyone 2 is in its shape-agnostic mask strategy, which more effectively characterizes the relationship between characters and their environments. Then, to improve the fidelity of object interactions, the model employs an object-guider to extract relevant features from interacting objects and utilizes spatial blending for feature injection. Additionally, a pose modulation strategy is introduced, allowing the model to accommodate a wider range of motion patterns.

Experimental results indicate that Animate Anyone 2 significantly outperforms its predecessor. And finally, Perplexity just introduced Deep Research, a groundbreaking tool designed to streamline in-depth research and analysis. This innovative feature autonomously performs extensive searches, reviews hundreds of sources, and synthesizes findings into comprehensive reports within minutes. Available for free to all users, Deep Research offers pro subscribers unlimited queries, while non-subscribers can access a limited number daily. Now tell us if you think this is going too far. [tr:trw].