Are you stressed out trying to keep up with rapid-fire developments in humanoid robotics? You’re right where you should be!
To keep you in the know, here’s the latest in humanoid robots, including a preview of what Figure calls the most advanced humanoid robot ever, NVIDIA tools for accelerating deployment, the push for physical AI, and a budget-friendly option for DIY-ers and more updates you probably missed if you blinked.
Subscribe to my YouTube channel to stay up to date as humanoid robotic technology advances rapidly. Leave a comment to suggest humanoid robots to cover in future roundups.
Figure 02 Preview
Figure 02 is the most advanced humanoid robot on the planet, according to the CEO of Figure, the OpenAI startup whose humanoid robot is now working at a BMW factory.
Brett Adcock announced on X that Figure 02, which follows Figure 01, will be unveiled on August 6th. Figure released a teaser that VentureBeat notes is short on specifics but heavy on vibes.
The Figure 01 humanoid robot recently started working at the BMW plant in Spartanburg, South Carolina. Figure describes itself as a first-of-its-kind AI robotics company bringing a general-purpose humanoid robot to life.
Earlier this year, Figure announced it had secured $675 million in funding from OpenAI and others. The partnership accelerated the development of Figure 01’s conversational abilities. Weeks after the February announcement, Figure released footage of the robot conversing in real-time.
NVIDIA’s Tools for Accelerating Deployment
NVIDIA has released new tools aimed at accelerating the deployment of humanoid robots worldwide.
The tech giant just announced a suite of tools for robot simulation, learning, and training. A new humanoid robot developer program provides early access to the tools, with companies like Boston Dynamics and ByteDance already participating.
NVIDIA says its NIM microservices can reduce robot deployment times from weeks to minutes. The MimicGen NIM microservice creates synthetic motion data, while the RoboCasa NIM microservice makes simulation-ready environments. Osmo, NVIDIA’s cloud-based service, simplifies robot training and simulation. The company says it cuts development times from months to under a week.
Teleoperation, which captures human actions to train robots, can be done faster and cheaper with NVIDIA’s AI and Omniverse digital twin tools.
The software is compatible with the Apple Vision Pro, which captures human demonstrations that are simulated and expanded into large datasets for training robots.
NVIDIA has been vocal about humanoid robots, investing heavily in their development and capabilities. Program members get early access to foundational models in NVIDIA’s Project Group. The project, announced earlier in 2024, uses NVIDIA’s tech to teach robots how to move and interpret human language.
A robotic startup could use a foundational model with their hardware as a starting point. Since the model already knows a lot, the startup can focus on teaching the robot to perform tasks that can’t yet be automated.
AI Shopping Companion
MenteeBot is an AI shopping companion in robotic form.
Mentee Robotics says a large international retailer suggested deploying robots to push shopping carts for the elderly. The Israeli company says it was also approached by a nursing home chain considering robotics to remedy labor shortages. In response, the team quickly developed a robot that could grab a shopping cart and follow someone in a wheelchair.
The robot stops when its shopping companion does and pushes the cart smoothly.
The robot is expected to be available to select customers in 2025.
RialTo: Digital Twins for Training Robots
Robot success increases by over 67% using digital twins of real-world locations for training.
That’s according to researchers from the US and Germany who developed a new method for training robots with computer simulations called real-to-sim-to-real or RialTo.
- First, the physical location is scanned to create a detailed digital copy.
- With that, a digital twin is built that reflects the physical space in real-time.
- The robot software operates in the virtual environment, letting it see and interact with objects as if they were real. Robots practice tasks in the digital twin, safely learning and improving their skills virtually.
- After training in the simulation, the research team says the robots perform the tasks in the real world with much greater success.
The study used Franka Emika Panda robotic arms. Tasks performed included stacking dishes and placing books on a shelf. The researchers are affiliated with the Massachusetts Institute of Technology, the University of Washington, and the Technical University of Darmstadt. The study was partly funded by the Sony Research Award Program, the US federal government, and Hyundai.
RX1: Budget-Friendly DIY Humanoid Robot
The RX1 humanoid robot is open source and can be built for under $1,000.
That’s according to Lingkang Zhang, founder of Red Rabbit Robotics in British Columbia, Canada. Zhang says the RX1 has full human-scale robotic arms, built-in stereo vision, and understands the world in three dimensions. It can be trained to complete tasks using simulation software.
Skild AI’s Foundational AI Model
Skilled AI has secured $300 million to build a foundational AI model to power robots for real-world applications. The round of Series A funding brings the San Francisco-based startup’s valuation to $1.5 billion.
The company says the funding will go toward advancing AI to make robots more adaptable and useful across industries. by Lightspeed Venture Partners, investors include Coatue, Softbank, Jeff Bezos, Sequoia, Felicis Ventures, and Amazon.
Skild AI says its foundational model is trained on a significantly larger and more diverse set of data compared to competitors. The company says it’s working on hiring experts in AI, robotics, engineering, operations, and security to help bring advanced AI into physical reality.
Elephant Robotics’ Mercury X1 and B1
Elephant Robotics says its new humanoid robots help researchers bring artificial intelligence into the physical world. The Mercury X1 is a wheeled humanoid robot equipped with dual NVIDIA Jetson controllers, LiDAR and ultrasonic sensing, and an 8-hour battery.
The Mercury X1 is designed for mobile tasks and exploration. They also have a semi-humanoid robot, the Mercury B1, which is intended for advanced research. The robot features a 9-inch touchscreen, an NVIDIA Xavier chip, and a 3D camera.
The robots are geared toward researchers and educators focused on embodied AI. That’s when artificial intelligence is embedded into a machine, empowering it to move and interact with its physical surroundings autonomously.
The X1 is the first full humanoid for the Shenzhen, China-based Elephant Robotics. The company’s other robots include a series of cobots and robotic pets.
OpenTeleVision: Controlling Robots with VR Goggles
Researchers from UC San Diego and MIT have developed a new system for controlling robots with VR goggles from far distances. The system, called OpenTeleVision, provides an immersive experience.
The operator could be as far as 3,000 miles away, as demonstrated by an operator in Boston controlling a robot in San Diego.
It mirrors the operator’s hand and head movements on the robot, delivering immersive active visual feedback. The robot sends 3D video back to the operator for better visibility into the robot’s surroundings. Using the teleoperation method, robots can perform tasks like sorting cans, folding towels, and moving items.
The team tested the system with the Chinese-made humanoid robots Unitree H1 and Fourier GR-1. They used Apple Vision Pro and Meta Quest goggles for the experiments.
The researchers reported high success rates. The H1 and GR1 had a 92% and 87% success rate for picking up cans, respectively. They had more trouble placing the cans in the correct position, with the H1’s success rate at 88% and the GR1’s at 60%. However, they both had a 100% success rate folding towels.
In its report, the researchers said operators found it easier to control robots using this system compared to traditional methods. The system’s code and setup are open source and available online for anyone to use.
The Latest in Humanoid Robotics
For the latest in humanoid robots, subscribe to my YouTube channel.