Character Founder and CEO Brett Adcock Thursday reveal A new machine learning model for humanoid robots. The news comes two weeks after Adcock announced that Bay Area Robots will arrive Decided to leave OpenAI cooperationHelix, a model of “generalized” visual language (VLA).
VLA is a new phenomenon used in robotics, using visual and linguistic commands to process information. Currently, the most famous example of this category is Google DeepMind’s RT-2train the robot through a combination of video and large language models (LLM).
Helix works in a similar way, combining visual data and language to prompt real-time control of the robot. “Helix shows a powerful object generalization, able to thousands of novel household items never encountered in training in different shapes, sizes, colors and material properties, just inquiry in natural language,” the picture reads.
In an ideal world, you can simply tell the robot to do something, and that will only do it. According to the numbers, this is where the spiral comes in. The platform is designed to bridge the gap between visual and language processing. After receiving a natural language prompt, the robot visually evaluates its environment and then performs a task.
The figure provides examples such as “handling the cookie bag to the robot on the right” or “receiving the cookie bag of the robot on the left and then placing it in an open drawer”. “Both examples involve a pair of robots working together. This is because Helix aims to control two robots at once, and another assists another robot in various home tasks.
The figure shows VLM by highlighting the work done by companies using their Class 02 human robots in a home environment. Houses are known to be tricky for robots because they lack the structure and consistency of warehouses and factories.
Difficulties in learning and control are the main obstacles to standing between complex robotic systems and houses. These issues and the five to six-digit price tags, which is why home robots don’t prioritize most humanoid robotics companies. Generally speaking, this approach is to build robots for industrial customers, which can both improve reliability and reduce costs before addressing the home. Housework is a conversation a few years from now on.
When TechCrunch Touring People Bay Area Office In 2024, Adcock showed off the pace of the company placing its humanoid organisms in a home environment. At the time, the work was not prioritized because the numbers focused on workplace pilots with companies like BMW.
With Thursday’s spiral announcement, the numbers make it clear that the home itself should be a priority. This is a challenging and complex environment for testing such training models. For example, robots are taught to perform complex tasks in the kitchen, for example, in different environments, they turn them on various actions.
“To make robots useful in the home, they will need new behaviors that are able to generate intelligent on demand, especially for objects that have never been seen before,” Figie said. “Teaching robots, even a new behavior, currently requires a lot of human effort: hours of doctoral specialists manually programming or thousands of demonstrations.”
Manual programming does not extend for the house. There are too many unknowns. The kitchen, living room and bathroom vary greatly from one to another. The same can be said for tools for cooking and cleaning. Additionally, people leave behind chaos, rearrange furniture, and love a range of different ambient lighting. This approach takes too much time and money – despite the numbers Of course there are many latter.
Another option is training – a lot. This method is often used to trained robot arms pick and place objects in the lab. What you can’t see is the hundreds of hours of repetition that require enough demonstration to take on highly variable tasks. To pick up something for the first time, it used to be a robot that took hundreds of times.
Just like the humanoid robotics around it, working in the spiral is still at a very early stage. It should be advised that many work happens behind the scenes to create the short, well-made videos seen in this article. Essentially, today’s announcement is a recruitment tool designed to bring more engineers on board to help grow the project.
