The robot can perform tasks completely independently without human remote control or preset scripts.
All movements are calculated in real time through a neural network.
The robot’s vision-based end-to-end neural network learns directly from images how to control its movements, including driving, manipulating arms and graspers, and controlling the trunk and head.
The presentation video was not edited at an accelerated pace…
Through training, robots are able to understand and perform a wide range of physical behaviors, such as cleaning, organizing, picking objects, and socially interacting with humans and other robots.
The project uses a strategy that allows the robot to quickly learn new skills by collecting data in minutes and training it on a desktop GPU to quickly fine-tune the model to suit specific tasks.
Technical principle:
1. Demonstration data set construction: The team first collected a set of demonstration data containing 30 EVE robots performing various tasks. The data is very diverse and includes examples of robots performing physical behaviors such as cleaning, organizing homes, picking up objects, and interacting socially with people and other robots.
2. Basic model training: Using these demonstration data, the team trained a “basic model”. This model is able to understand a wide range of physical behaviors and provides a basis for robots to understand the basic actions of various tasks.
3. Model fine-tuning: Next, the team fine-tuned the basic model for specific task types and generated multiple specialized capability models. For example, they created a model dedicated to door operations and another model focused on warehouse tasks.
4. Further fine-tuning to suit specific tasks: The team further fine-tuning these specialized models to enable them to perform more specific tasks, such as opening a specific door. This fine-tuning process allows the model to have a more precise understanding and execution of specific behaviors or actions.
5. Rapidly introduce new skills: With this hierarchical fine-tuning strategy, teams can quickly introduce new skills to the robot in a short period of time (just minutes of data collection and training on desktop GPUs).
Details:https://1x.tech/discover/all-neural-networks-all-autonomous-all-1x-speed
There are no remote controls, no computer graphics, no clips, no video acceleration, and no script track playback in the demonstration video. Everything is controlled through a neural network, completely autonomous, and proceeds at 1X speed.
Robot hardware information:
The goal of 1X is to design a universal Android robot that can work effectively in any scenario to cope with the unpredictability of the real world.
EVE is an advanced Android robot developed by 1X Technology to provide smart work solutions for the business industry. EVE is designed to perform a variety of tasks, from logistics operations to security patrols. This robot combines safety, balance and intelligence, and can be easily integrated into existing workflows and seamlessly collaborate with teams.
The main features include:
Safety first: Each EVE is tested in real-world scenarios before deployment. Its soft, biomimetic mechanical design ensures safety from the inside out and is suitable for working in a variety of spaces.
Balancing performance: EVE is capable of handling heavy objects while also being gentle enough to handle fragile items, making it easy to integrate into your logistics workflow whether in a warehouse or a distribution center.
Intelligent behavior: EVE learns by observing the expert’s actions, such as moving the device, opening the door, completing an order, etc., and then reproduces the actions through the learning model it embodies and starts working. Over time, EVE builds experience on basics, making future tasks, such as “moving that box”, easier.
Specifications of EVE:
Height: 1.86 meters
Weight: 86 kg
Maximum speed: 14.4 km/h
Carrying capacity: 15 kg
Running time: 6 hours
EVE operates autonomously through artificial intelligence and is able to navigate the workspace by default and perform tasks such as opening doors with different handles, identifying people or objects from a distance, and traveling through unstructured space like humans. While helping human operators perform tasks, EVE uses its power, accuracy and sensors to perform tasks such as patrolling office buildings and checking employee ID badges. EVE will focus on potential hazards or errors and report them when the operator needs to take over.
Video: