NEW $8999 Open Source K-Scale Humanoid Robot AI Roadmap 2025-2028 (CL-3 UPDATE)

➡ A new AI robot called K-Bot, priced at $9,000, is set to launch in November. It’s designed to learn, adapt, and connect remotely, and can perform everyday tasks due to its compact size and long battery life. Another robot, the CL3 by Linux Dynamics, can mimic human movements and navigate complex environments. Meanwhile, AI researchers from Carnegie Mellon University have developed DexWild, a system that captures human hand movements to improve robot interactions, potentially leading to more reliable and adaptable robots in various environments.

Is this the most affordable humanoid yet? A new $9,000 AI robot named the K-Bot is preparing for its release this November as the very first open-source, full-stack personal robot designed for adaptability, learning and remote connectivity. Standing 4 feet 7 inches tall and weighing 77 pounds, the K-Bot combines a compact frame with a maximum payload capacity of 22 pounds, making it suitable for everyday tasks across a wide range of environments. But one of the K-Bot’s key strengths is in its runtime, offering up to 4 hours of battery life per charge. And when it comes to visual awareness, the K-Bot supports a suite of cameras, including RGB, stereo vision and an ultra-wide field of view, while far-field microphones and stereo speakers can enable natural voice interaction.

On top of this, its reinforcement learning is backed by millions of hours of simulation, allowing K-Bot to withstand external physical forces as well as enhance its ability to navigate complex terrain and handle dexterous manipulation tasks. And when it comes to hardware, the K-Bot is modularly designed for extensibility, allowing users to upgrade sensors or compute power as technology advances, without needing to discard the entire system. For dexterity, the adaptable end-effector design means the robot can be equipped for a variety of custom applications for hobbyists, researchers and industry use cases alike.

And connectivity is central, as the K-Bot is internet-enabled to be managed and controlled from anywhere, meaning that telepresence is also an option. At the heart of it all is the KOS Robot Operating System, developed in Rust, which handles both low-level reinforcement learning locomotion with System 1 and high-level foundational tasks utilizing integrated vision and language models with System 2. And looking ahead, the company has outlined its roadmap, with an initial release planned in November for basic locomotion, balance, and app-based control. Then this December, the company plans to bring vision-language action integration to natural language tasks, with full autonomy being planned for 2028, requiring less than one human intervention per day.

Altogether, K-Bot’s open-source approach is geared towards developers and end-users alike. At the same time, there’s another brand new humanoid robot from Linux Dynamics that’s approaching its commercial release too, but how smart is it, and what can it do? Standing at approximately 5 feet 5 inches, or 164 centimeters tall, the CL3 has already proven it can replicate a series of human-like movements in previous videos. In the latest footage however, the robot is seen walking down the street in China, while previous videos showcased the robot lifting its arms overhead and twisting at the moves.

The overall effect is impressively lifelike as Linux Dynamics demonstrates the latest progress in robotic motion control, with the CL3’s technical specifications being key to these advances. The robot’s design includes between 29 and 52 degrees of freedom, allowing for highly articulated body movements. Plus, Linux Dynamics employs a proprietary hollow actuator design with high torque density actuators in order to deliver both the strength and fluidity needed for such complex full-body movements. Furthermore, the company emphasizes that these features enable ultra-stable and dynamic movement, even when performing coordinated routines. On top of this, the CL3 is outfitted with advanced onboard sensors and artificial intelligence algorithms.

These systems allow the robot to perceive terrain in real time to adjust its balance and gait accordingly, making it possible for the robot’s autonomous navigation across complex environments. Currently, not many other details about the robot have been released, so its level of dexterity, speed, computational power, and other key metrics are still to be discovered with the company’s upcoming release. And for next-level robot dexterity, AI researchers from Carnegie Mellon University also just revealed a breakthrough system called DexWild, which is a comprehensive approach to advanced dexterous human-robot interactions far beyond their current limitations, but now with a twist.

That’s because at the heart of DexWild is a portable, high-fidelity, and embodiment agnostic data collection system that captures dexterous human hand movements in a wide range of indoor and outdoor environments. In fact, the DexWild initiative has already amassed an impressive 9,290 human demonstrations spanning 93 distinct environments, all collected at 4.6 times the speed of traditional teleoperation-based data gathering, which not only accelerates progress but also dramatically reduces costs and logistical challenges. Furthermore, DexWild stands out due to its special ability to align both the observation and action spaces between humans and robots.

By leveraging palm-mounted cameras, delivering a broad field of view focused on the task and the surrounding environment, while minimizing embodiment-specific bias, the team can co-train data from various robot forms and human demonstrators. All hand poses are systematically converted into a standardized 17-dimensional joint angle format, and wrist movements are mapped into a six-dimensional action space, enabling seamless policy learning across differing robot architectures. On top of this, DexWild demonstrates remarkable prowess in what researchers call ex-embodiment transfer. When policies are trained with DexWild data, robots are capable of zero-shot transfer to completely new robot arms and few-shot adaptation to novel robot hands.

This flexibility is possible because DexWild training data is embodiment-agnostic. The captured information is solely task-relevant, allowing the policies to generalize and retarget between multiple robot types. For instance, policies trained on Leap2Advanced and Xarm6 platforms have been successfully deployed on FrancaPanda arms and LeapHands, proving DexWild’s broad compatibility and adaptability. But what’s especially noteworthy is how DexWild trained policies outperform traditional robot-only training by a significant margin, delivering four times higher success rates in unfamiliar environments and 5.8 times better transfer across different robot embodiments. Plus, these policies generalize robustly to entirely unseen settings, handling crowded, cluttered, and variably lit scenarios with ease.

The diversity and quality of the DexWild dataset are key contributors to these advancements, as evidenced by successful skill transfer between tasks, such as moving from spraying to pouring, without necessitating new robot data. Importantly, the research team highlights that scaling robot-only datasets is both expensive and restrictive, often yielding solutions that overfit to specific robots or environments. In contrast, integrating diverse human demonstration data through DexWild introduces more challenging learning conditions, ultimately producing policies that are more robust and capable of adapting to real-world complexity. With DexWild, the field of robotics takes a significant leap forward, potentially paving the way for more reliable, adaptable, and intelligent robots in everyday environments, whether at home, in industry, or navigating the unpredictable outdoors.

But what makes DexWild even more fascinating is its potential to democratize robotic learning. Because its data collection system is lightweight and portable, researchers and developers across the globe, regardless of lab funding or geographic location, can now contribute to and benefit from a shared repository of high-quality, real-world hand manipulation data. This opens the door to a globally distributed training model for dexterous robotics, accelerating collaborative progress in the field like never before. Anyways, make sure to like and subscribe for more AI news. [tr:trw].