Engine AIs $25500 PM01 Humanoid Robot Debuts New 23 Axes Tech (GPT5)

➡ Engine AI has unveiled its new humanoid robot, PM01, with human-like features and advanced mobility. It uses a combination of reinforcement and imitation learning techniques for natural movement. In other AI news, Skilled Brain has introduced a vision-based model for real-world locomotion, Astrobot has announced breakthroughs in full-body mobile operation, and Google DeepMind has released Genie 3, a simulator that creates interactive environments. Lastly, OpenAI has launched GPT-5, a unified system for adaptive reasoning and complex problem solving.

Today on AI News, new robot features and pricing, interactive world models, and GPT-5. But first, Engine AI just released its newest PM01 humanoid robot demonstration, showcasing a range of human-like features, plus its updated price for two separate versions. So how much do they cost now? Before we get into that, starting with specs, the robot is 138 centimeters tall and weighs 40 kilograms, including 23 degrees of freedom in total, featuring five in each arm, and one in its neck. On top of this, PM01’s waist can rotate up to 320 degrees, delivering a maximum joint torque of 145 newton meters, for more fluid movement across a variety of environments.

But the PM01 is also designed for next-gen locomotion, employing reinforcement and imitation learning techniques paired with its proprietary harmonic joint modules. This combination allows the robot to walk with a remarkably natural gait, achieving speeds of up to two meters per second. Mobility is further enhanced by six degrees of freedom in the legs, three in the hips, one in the knee, and two in the ankles, contributing to stability and adaptability on uneven surfaces. And for computational power, Engine AI has equipped the PM01 with a high-performance dual-chip setup, combining the Intel N97 processor with NVIDIA’s Jetson Oren.

This architecture supports end-to-end neural network models for improved task execution and decision making. As for perception, the robot features a built-in depth camera to enhance its visual understanding of the environment, while a multi-array microphone and surround sound three-dimensional speakers support robust multimodal interactions. And powering it all is a 10,000 milliamp-hour quick-release battery, enabling approximately two hours of operation on a single charge and the option for rapid battery swaps to minimize downtime. To protect its internals, the robot is engineered with a full aluminum exoskeleton featuring integrated sealing and heat dissipation for maximum durability.

As for pricing, Engine AI has listed the commercial version at $25,500, while the educational version is set at $39,500, with both models being open source, designed to serve in both home and work environments. Simultaneously, to make fully generalizable robots a reality, the team behind Skilled Brain just introduced their newest vision-based model, designed to tackle real-world locomotion in a way that mimics human adaptability. Unlike previous systems that relied on detailed mapping or preplanned behaviors, the Skilled Brain enables a robot to interpret raw camera images and joint feedback, outputting precise motor commands for smooth, instinctive movement.

Impressively, the robot navigates city stairs, fire escapes, and unpredictable terrain entirely on the fly, with no manual switching between walking modes required. Furthermore, the team tested the system on tricky obstacle courses with unstable pallets and varying step heights, with the model adjusting in real time. In fact, the robot even managed to carry boxes up and down stairs using its single end-to-end neural network, with even broader applications planned for release in the near future. But Astrobot also just announced two new breakthroughs in the area of general robot intelligence. First, the company introduced DuoCore WB, its first full-body mobile operation model, which enables robots to advance from half-body to full-body intelligence.

And Astrobot isn’t restricted to static or tabletop tasks, instead autonomously adapting manipulation strategies based on its environment, with the model integrating sensory input with motion control. Next, the robot’s real-time trajectory generation module allows for smoother, more stable operation by tightly connecting with the underlying control system. And to accelerate real-world adoption, Astrobot also rolled out the Astrobot suite as a comprehensive learning platform to make it easier for robots to acquire and master complex daily skills even faster. And as robots adapt to new environments, Google DeepMind just taught AI to create, understand, and interact with new worlds with its release of Genie 3.

Designed as a general-purpose simulator, Genie 3 generates a remarkable variety of interactive environments from simple text prompts, allowing users to navigate dynamic AI-created worlds in real time at 24 frames per second and at a resolution of 720 pixels. Importantly, world models allow AI to simulate their environments so that agents can predict how the world will evolve and how their actions can change the course of those environments. In fact, these models are an important stepping stone on the path towards artificial general intelligence, making it possible to train agents in endlessly diverse and evolving situations.

Previous iterations, such as Genie 1 and Genie 2, introduced foundational world models that could generate new environments for agents, while other DeepMind models like VO2 and VO3 demonstrated state-of-the-art video generation and a deep understanding of intuitive physics. But Genie 3 sets itself apart by offering real-time interactivity and improved consistency over its predecessors. The model supports a wide range of capabilities, including the simulation of physical properties such as water, lighting, and dynamic environmental interactions. Users can even experience natural phenomena as well as vibrant ecosystems, complete with complex animal behaviors and intricate plant life.

On top of this, Genie 3 enables the creation of imaginative, animated scenarios, expressive characters, and facilitates the exploration of historical settings and distant locations. But achieving real-time interactivity required DeepMind researchers to make key technical breakthroughs. Amazingly, for each frame, Genie 3 references a growing trajectory of previous interactions, allowing for consistent responses even if a user revisits a virtual location after a minute has passed. This ability is crucial for immersive simulations, where environmental consistency over extended periods is essential. Notably, Genie 3’s environments remain visually and physically consistent for several minutes at a time, with a visual memory that can extend as far back as one minute, an emergent capability that distinguishes it from autoregressive video generation techniques where inaccuracies can quickly accumulate.

Another major advancement in Genie 3 is the introduction of promptable world events. Beyond traditional navigation controls, users can now employ expressive text-based interactions to trigger changes in the environment, such as shifting weather conditions or introducing new objects and characters. This added layer of control expands the potential for counterfactual scenarios, giving artificial intelligence agents richer and more varied experiences to learn from, especially when preparing for unexpected real-world situations. And finally, OpenAI has officially launched GPT-5 with a unified system designed for adaptive reasoning and complex problem solving. Curiously, GPT-5 merges all prior model lines into a single flexible architecture that dynamically adjusts its reasoning depth to match the complexity of each task.

Everyday queries are handled by GPT-5 main, while more challenging problems are routed to GPT-5 thinking. Next, an intelligent router chooses the best model for each question based on difficulty, conversation context, or direct user prompts, with pro subscribers having access to GPT-5 pro, which dedicates even more computation to the toughest queries. Furthermore, GPT-5 introduces improvements across key benchmarks, including increased accuracy in coding while slashing error rates by two-thirds compared to O3. In healthcare, GPT-5 scores 14.6% higher on healthbench hard, and in math and multimodal tests, GPT-5 pro reaches 88.4% on the GPQA science benchmark, but reducing factual errors remains a core focus.

With web search enabled, GPT-5 is about 45% less likely to make mistakes than GPT-40, and in pure reasoning mode, the error rate drops by 80% compared to O3. As for safety, OpenAI has also implemented even more controls in its responses to prevent GPT-5 from offending anyone with the truth. But tell us in the comments what you think about GPT-5, and check out this video here for more AI news. [tr:trw].