What impact could OpenAI’s new Strawberry model have on robotics?
The newly released model, also known as o1, helps computer solve problems step by step. It uses Chain-of-Thought reasoning, which means it breaks big problems into smaller parts to find the best solution. The AI tool solves complex problems using advanced math and logic.
Unlike earlier models, ChatGPT o1 focuses on solving tough math problems and debugging code, not just writing text. Several high-profile humanoid robots like the Figure 02 and Sanctuary’s Phoenix use OpenAI. Advancements in ChatGPT could improve their ability to perform tasks and interact smoothly
It is better at handling multi-step tasks and deeper reasoning, which is useful in robotics. This could help robots understand physics better, allowing them to move and interact more naturally in the real world.
The model can help robots fix themselves when something goes wrong and collaborate smoothly in groups. With improved reasoning skills, robots could act in ways that seem more human-like, adapting to their environment more intelligently.
1X World Model
1X Technologies says it’s taking a radically different approach to training its humanoid robots for the real world.
The OpenAI-backed startup is developing a so-called world model that predicts outcomes when robots interact with things like doors, boxes, or even people.
The model could help robots better perform household tasks like folding clothes, opening doors, picking items up, and moving around without bumping into things. It helps AI-powered robots improve by showing different possible results from their actions.
1X says the world model beats traditional simulation-based training because it learns from real-life data. The data comes from thousands of hours of robot data collected during tasks performed from 1X robots in settings like homes and offices.
In a blog post, 1X said physics-based simulations like Bullet, Mujoco, Isaac Sim, and Drake are useful for testing robot actions because they can be reset and repeated to compare strategies. However, they’re mainly built for rigid objects and require manual configuration.
They struggle to simulate tasks like opening boxes, cutting fruit, and interacting with humans. So they don’t accurately reflect the complexity and variety of the real world especially for large-scale testing.
The method has limitations that the Norweigian company is trying to solve with its 1X World Model Challenge where people can win cash prizes for improving the world model:
- The model sometimes fails to maintain the shape or color of objects during interactions.
- While the model has some understanding of physical properties, it occasionally generates outcomes that defy real-world physics like floating objects.
- The model also struggles with self recognition.
Boston Dynamics’ Spot Tested at Fusion Facility
In a world first, the UK Atomic Energy Authority and Oxford Robotics Institute just trialed Boston Dynamics’ Spot quadruped robot to inspect inside a fusion energy facility.
The AI robot, controlled by the institute’s AutoInspect platform, worked in the Joint European Torus, a large research machine, for 35 days straight.
Humans can’t survive the conditions due to radiation, high temperatures, and pressure.
The robot mapped the facility, collected data, and avoided obstacles while operating without human control.
The test showed robots can make fusion energy maintenance safer and cheaper, and could be used in more dangerous environments like nuclear decommissioning.
VersaBot VB-1 Pure Vision-Based Humanoid Robot
Lanxin Robotics just introduced what it calls the world’s first pure vision-based humanoid robot.
The VersaBot VB-1 robot mimics human vision to achieve high-level spatial awareness. The humanoid robot uses cameras to create three dimensional maps of its surroundings.
The Chinese AI startup says VersaBot has a 360-degree view to detect and avoid obstacles. It uses an advanced RGB-D camera that captures color and depth.
With its powerful computing, the robot decides in real-time whether to stop or move around obstacles. The robot is primarily intended for industrial settings.
Lanxin says pure vision systems are more cost-effective and simpler to produce than machines that use multiple sensors for navigation.
Pudu Robotics D7 Robot
Pudu Robotics just unveiled its PUDU D7 semi humanoid robot that can press elevator buttons, choose the correct floor, and ride an elevator on its own.
The AI robot has a human-like upper body with bionic arms. It can move in all directions with its omnidirectional chassis. It’s designed for hotels and hospitals to deliver items across floors using elevators autonomously.
China-based Pudu says the robot can lift up to 10 kilograms. It works continuously for over 8 hours.
Pudu aims to launch the D7 for commercial use in 2025.
Robot Chefs Approved in Beijing
Robot chefs are coming.
Beijing just issued the first-ever license for the use of robots to cook and serve food. Chinese robotics company EncoSmart can now deploy its Lava robots to prepare fried foods.
The robots are fully autonomous. CEO Chen Zhen told the South China Morning Post the data collected from restaurants will help improve the robots over time.
Founded in 2022, the startup raised about $5.6 million to support its work on robotic chefs. The robots can cook and serve a batch of fries in about two minutes.
They learn to cook new dishes autonomously, adapting to different recipes. With special cameras, they recognize ingredients and adjust cooking times.
Google DeepMind DemoStart Training Method
Google DeepMind says its new method trains robots 100 times faster than older methods.
What would normally take 27 hours of real-world demonstrations now takes just 30 minutes in simulation.
The DemoStart method uses computer simulations to teach robots how to move and handle objects. The robots only need a few examples to learn quickly and perform perfectly in the real world. The method creates an automatic learning path that gets harder as the robot learns, helping it get better at tasks.
Even with limited, imperfect examples, DemoStart beats older methods. Robots learned tasks like inserting plugs, turning cubes, and threading bolts with high success.
Robots use cameras and touch to perform tasks in both virtual and real worlds.
More AI Robot Updates
To stay up to date on the most advanced AI-powered robotics, subscribe to my YouTube channel.