AGI Bot Humanoid Robot Brain Learns 1000000+ Skills NEVER Taught (AI NEWS)

In this model, human teleoperators perform tasks dynamically, responding to real-time conditions rather than following predetermined sequences. And it results in data that carries significantly more environmental variability and task complexity than what controlled lab settings typically produce. Furthermore, the dataset is built on AGIbot’s G2 hardware platform, which integrates high-performance joint actuators and multimodal sensors. This means the system captures synchronized data streams, including RGB imagery, tactile signals, lidar point clouds, inertial measurement unit readings, and full-body joint states, all within a unified pipeline, with each data episode undergoing both cleaning and validation through a multi-stage processing system before being released.

Plus, another notable feature is how the data is labeled. Each task is broken down into layered instructions, and most importantly, the dataset includes moments where things go wrong, not just the successful demos. This means a robot can now learn not only how to complete a task, but also how to recover when it makes a mistake. And AGIbot is planning to release the datasets in five phases, starting with imitation learning. On top of this, AGIbot is also releasing one-to-one digital twin simulation alongside every real-world episode, which bridges the gap between how robots train digitally and how they perform in the real world.

In fact, the dataset is already available through Hugging Face and includes over one million trajectories from 100 robots across more than 100 real-world scenarios and five target domains. But collecting and releasing massive amounts of real-world data addresses only half the equation. The other half lives inside a simulation platform, and that platform is called GenieSim 3.0, and it may represent a fundamental shift in how the robotics industry thinks about simulation. So, rather than treating simulation as a secondary tool that’s used to supplement real-world testing, AGIbot built GenieSim 3.0 as a comprehensive development environment to integrate scene generation, data collection, training, and evaluation into a single unified workflow.

And it’s built on NVIDIA’s Omniverse, its fully open source, plus its standout capability is its use of large language models to drive scene generation. This lets users describe environments in natural language so that Genie 3.0 can automatically produce structured three-dimensional scenes, visual previews, and thousands of semantic variations, all without requiring any manual coding. Then, vision language models can refine the scenes to meet specific requirements, meaning that what previously took hours of manual environment construction can now be accomplished in just minutes. And on the data side, GenieSim 3.0 includes what AGIbot describes as the largest open source simulation dataset in embodied AI, more than 10,000 hours of synthetic data covering over 200 tasks with multi-sensor modalities.

The platform also provides an automated data collection toolkit that supports both teleoperation and scripted task execution, with built-in auto annotation and a recovery mechanism that resumes data collection after task failures. Furthermore, GenieSim 3.0 also integrates a reinforcement learning pipeline, allowing robots to practice and refine precise skills entirely within simulation before ever touching a real object. And for evaluation, the platform offers over 100,000 simulation scenarios and uses AI to automatically generate testing workflows. AGIbot even reports that the gap between simulation and real-world tests results in less than 10%, and that in some cases, models trained purely on synthetic data have outperformed those trained on real-world data.

Plus, the platform also reconstructs one-to-one digital twins of real industrial environments, and all simulation assets, datasets, and evaluation source code have already been released as open source on GitHub. But a dataset and a simulation platform are only as useful as the models that they learn from. That’s where GoTo comes in, AGIbot’s next-generation foundation model for embodied AI. It was announced just yesterday, and its predecessor, Go1, gave robots the ability to understand instructions and plan tasks. But as those robots moved into more complex real-world environments, a problem emerged. They could only come up with reasonable plans, but couldn’t always execute them precisely.

So now, GoTo is designed to solve this, unifying reasoning and physical execution within a single architecture. Trained on tens of thousands of hours of data, it represents the final piece of AGIbot’s stack. The dataset provides the learning material, GenieSim 3.0 provides the training and testing environment, and GoTo is the model that turns all of it into real-world action. But while AGIbot is working to give robots the data and simulation tools they need to operate in the physical world, Anthropic is tackling a parallel challenge in the digital one, making AI agents that can actually ship production-ready work.

To accomplish this, Anthropic just launched Claude Managed Agents. Set of APIs designed to let developers build and deploy cloud-hosted AI agents at scale. And the core pitch here is speed, as what previously took months of infrastructure work can now be done in days. For instance, building a production-grade AI agent typically requires sandbox code execution, state management, credential handling, scope permissions, and end-to-end tracing. But now, Anthropic’s managed agents can handle all of that behind the scenes. Developers simply define the agent’s tasks, tools, and guardrails, and Anthropic’s infrastructure runs the rest, including deciding when to call the tools, how to manage context, and how to recover from errors.

Furthermore, the platform supports long-running sessions that can operate autonomously for hours and even persist through disconnections. Plus, there’s also a multi-agent coordination feature that’s currently in research preview, where agents can spin up and direct other agents to parallelize complex work. And several companies are already using the platform. Notion integrated it to let its teams delegate tasks like coding, building presentations, and generating spreadsheets directly within their workspace. Rakuten deployed specialist agents across engineering, product, sales, and finance, each within a week, century paired it with their debugging tools so that developers can go from a flag bug to a reviewable code fix in a single workflow.

Asana used it to accelerate the development of what it calls AI teammates or agents that work alongside humans inside project workflows, and in internal testing, Anthropic reported that managed agents improve task success by up to 10 points over standard prompting, with the largest gains on the most difficult problems. So for now, the platform is available in public beta, priced on consumption at standard token rates plus eight cents per session hour for active runtime. Anyways, like and subscribe for more of the latest in AI and robotics news, and thanks for watching.