New Walker S2 AI Robot Breakdown RobotEra L7 AI News $45000 Humanoid Robot

➡ Uptech has launched a new AI system that enhances the capabilities of its Walker S2 robot, allowing it to interact with its environment like a human. Meanwhile, RobotEra has introduced its L7 humanoid robot, which can replicate intricate hand motions and adapt quickly to changing spaces. PND Botics, Noitom Robotics, and Inspire Robots have partnered to offer a comprehensive data acquisition solution for robotics, while SEED has unveiled GR3, a large-scale vision-language action model that can adapt to new objects and environments. All these advancements are pushing the boundaries of AI and robotics, making them more versatile and adaptable for real-world applications.

Uptech just introduced its new dual-loop AI system, combining BrainNet 2.0 with a coagent structure to deliver advanced swarm intelligence for the real world. But what can it do now? To start, the new system mimics human binocular stereo vision, allowing the Walker S2 to perceive depth and interact with their environment much as a human would. On top of this, the robot is equipped with Generation 4 industrial dextrous hands, offering 11 degrees of freedom and a 6-tactile sensor array, enabling each hand to carry payloads of up to 7.5 kg, while individual fingers can manage objects weighing up to 1 kg.

For user interaction, the robot features a 4-inch circular facial display and a 4-microphone array with two output speakers. Furthermore, Walker S2 now uses a mix of rigid-flex heterogeneous composites, aerospace-grade aluminum alloy, 3D-printed components, and high-elastic fiber materials for enhanced strength and flexibility. All of these allow the robot to deep squat, stoop lift, and maintain stability even at a pitch angle of 170 degrees, with a workspace reach of up to 1.8 meters. Additionally, its new waist allows for a pitch angle of plus 90 to minus 35 degrees, and a rotation angle of plus or minus 162 degrees, allowing the robot to even autonomously charge and replace its own batteries within 3 minutes.

As for its size, the robot is 176 centimeters tall and weighs 73 kilograms. Despite its full size though, it still achieves a full speed of up to 2 meters per second, or just over 7 kilometers per hour, all supported by 52 degrees of freedom throughout its body. But competitors like RobotEra are turning up the heat with the debut of their newest L7 humanoid robot, showcasing the L7’s capabilities through not only a dance demonstration, but also a live sorting task to prove it. Standing at 171 centimeters tall and weighing 65 kilograms, the L7 is designed to closely match human proportions, with its full body flexibility featuring a total of 55 degrees of freedom.

Of these, the robot has 7 joints in each arm and 12 degrees of freedom per hand, while utilizing direct drive technology. This setup allows the L7 to replicate intricate hand motions, with a finger speed of 10 clicks per second, operating across a 2.1 meter reach and workspace. And powering its movements are RobotEra’s self-developed joint motors, which deliver up to 400 newton meters of torque, at 25 radians per second. With this tech, the L7 can carry a total payload of up to 20 kilograms when using both of its hands, achieving stronger and more dynamic operations. But the L7 is also notably agile, being able to jump and turn 360 degrees in the air, and boasts a record-breaking running speed of 4 meters per second.

For environmental awareness, the L7 is equipped with dual-camera vision and three-dimensional LIDAR, enabling 360-degree perception and rapid adaptation to changing spaces. Navigation and manipulation are guided by its ERA-42 vision-language action model, which coordinates all 55 degrees of freedom for both full-body and half-body configurations. Performance-wise, the L7 leverages a dual-processor setup. An X86 chip provides 80 Terra operations per second, while NVIDIA’s Jetson Oren delivers 275 Terra operations per second, supporting real-time processing for advanced tasks. Furthermore, the robot features two degrees of freedom in its neck, three in its waist, and six in each leg, ensuring balanced, human-like movement.

Overall, Robot ERA’s L7 combines high-level dexterity, environmental awareness, and next-generation computing power, suggesting the L7 is ready not just for show, but for practical work in the real world. Meanwhile, there’s a new player entering the robotics data acquisition field, as PND Botics partners with Noitom Robotics and Inspire Robots to unveil a comprehensive one-stop data acquisition solution package built around the ADAMU hardware platform, with its official launch planned for the World Artificial Intelligence Conference 2025. The robot’s pricing has been revealed to start at $45,000, but the new package actually integrates several advanced technologies, starting with Noitom’s PN-Link full-body wired Inertial Motion Capture Suit.

Next, Inspire Robots’ 6 degrees of freedom tactile dextrous hand delivers precise and realistic manipulation capabilities, working together with the ADAMU platform, which offers 31 degrees of freedom. This includes a 2 degrees of freedom head for nuanced positioning, 6 degrees of freedom dextrous hands, a 3 degrees of freedom waist, and a binocular vision system for advanced perception tasks. The waist mechanism also features a braking system, giving an added layer of safety and stability while allowing for flexible movement. Most of all though, the ADAMU’s rigorously optimized biomimetic structure is designed to enhance usability in data acquisition workflows, making it easier for users to collect and utilize human motion and manipulation data.

This is of particular importance for reinforcement learning and imitation learning applications, where robust and realistic datasets are crucial. Together, PNDbotics, Noitom Robotics, and Inspire Robots are providing an end-to-end ecosystem, supporting customers from initial research and development, all the way to production. With pre-orders now open, the ADAMU solution aims to bridge the gap between human movement capture and real-world robotic manipulation. And finally, a new chapter in generalizable AI robots is unfolding with the introduction of SEED GR3, a large-scale vision-language action model that promises to redefine how robots interact with the world. Designed for versatility and adaptability, GR3 demonstrates impressive capabilities in generalizing to new objects, environments, and instructions, particularly those involving abstract concepts.

Plus, its ability to be efficiently fine-tuned with only a small amount of human trajectory data means rapid and cost-effective adaptation to unfamiliar settings. At the heart of GR3’s generalizable performance is a multifaceted training regimen. The model is co-trained using web-scale vision-language datasets, which enhances its understanding of diverse scenarios and instructions. On top of this, it leverages user-authorized human trajectory data, collected via virtual reality devices, to fine-tune its performance for highly specific tasks. And complementing GR3 even more is the Bite Mini, a newly introduced bi-manual mobile robot engineered for flexibility and reliability. And when paired with GR3, Bite Mini can tackle a wide array of tasks, from precise pick-and-place operations to handling messy, real-life environments.

Extensive real-world experiments reveal GR3’s ability to reliably perform across three particularly challenging tasks. For example, in pick-and-place evaluations, GR3 was tested in four settings, familiar environments and objects, new environments, new instructions, and new objects, 45 of which were not seen during training. Results consistently show that co-training with large-scale vision-language data is crucial for strong generalization, as removing this step significantly decreases performance on novel tasks. Moreover, increasing the number of human-provided trajectories per object incrementally boosts its accuracy on unseen objects, while not affecting performance on previously seen ones. On top of these advances, GR3 proves its mettle in real-world, long-horizon manipulation tasks, such as autonomously cleaning a cluttered dining table or hanging clothes.

Anyways, like and subscribe for more AI news and check out this video here. [tr:trw].