Boston Dynamics and Toyota Collaborate: Training Atlas to Learn

The Toyota Research Institute has been making significant strides in teaching robots to quickly learn and execute tasks autonomously. Now, it is introducing its Large Behavior Model technology to the remarkable Atlas humanoid robot in collaboration with Boston Dynamics.

Surprisingly, humanoid robot hardware is already quite advanced. Over a decade of development at Boston Dynamics has produced not only the highly athletic and capable hydraulic Atlas robot but also a range of emerging commercial competitors, including Tesla, Figure, Agility, Sanctuary, and Fourier. While these impressive robotic bodies will continue to evolve, they are already proficient enough to carry out various useful tasks.

Software is the Key Challenge

The challenge is the software. Relying on a coding team for each new behavior provides minimal advantage over conventional production robots. Developing a general-purpose humanoid robot that can understand and interact with the world flexibly is a monumental task.

The solution lies in AI, which is becoming essential for many challenges but requires extensive training data. Models like ChatGPT, Grok, Llama, and Claude utilize the vast written data humanity has accumulated, enabling Large Language Models (LLMs) to process information quickly.

Overcoming Data Limitations with ‘Large Behavior Models’

In contrast, teaching robots basic movements suffers from a lack of data, as videos don’t fully explain the reasoning behind actions. Robots need to learn from scratch, leading to the concept of a ‘Large Behavior Model’ (LBM). This approach helps robots develop basic movements and combine them into complex actions, similar to how LLMs learn human language.

If you’re not familiar with the LBM research from the Toyota Research Institute (TRI) last year, it’s worth exploring.

Telepresence System Enhances Robot Control

The TRI team developed a telepresence system that enables human operators to control robotic arms using VR goggles connected to the robot’s cameras, offering a real-time view. Haptic gloves allow operators to feel sensations similar to those detected by the robot’s tactile sensors.

With the pilot limited to the robot’s sensory input, they performed various tasks in a kitchen, repeating them from different starting points, correcting mistakes, and evaluating each attempt.

Afterward, the robots “thought” about the tasks, simulating millions of variations and assessing their performance based on their understanding of success and failure.

Robotic Arms Master 60 Behaviors with Accelerated Training

By September of last year, the TRI team had taught the robotic arms over 60 complex behaviors. Researchers found that after an afternoon of piloted training, the system could run simulations overnight, allowing the robots to perform tasks independently by morning.

This remarkable work has great potential, and we look forward to seeing how much progress has been made in the past year, especially with rapid advancements in AI. However, it’s important to note that this research focused on pairs of robotic arms rather than full humanoid bodies.

That’s about to change. Boston Dynamics has long been the gold standard in robotics research, and the previous hydraulic Atlas humanoid is remembered as a groundbreaking machine in the field.

The New Electric Atlas

The new fully electric Atlas, which debuted just five months ago, sacrificed some of the original’s explosive power but compensates with swivel-capable joints throughout its body. This allows for rotation at the hips, shoulders, waist, neck, biceps, and thighs, enabling any part of its body to face any direction. From gymnast to contortionist… Check it out:

The new Atlas is an impressive robot, already displaying typical scratches and dents that give it a used look. However, in the past five months, we haven’t seen much of it to evaluate its capabilities, aside from its ability to do push-ups.

Today’s news is exciting: Boston Dynamics, a leader in humanoid robot hardware, is partnering with a top AI and Large Behavior Model (LBM) team to enhance the capabilities of humanoid robots.

Boston Dynamics and TRI Partner to Drive Humanoid Development

“There’s never been a more exciting time for robotics, and we’re eager to collaborate with TRI to accelerate the development of general-purpose humanoids,” said Boston Dynamics CEO Robert Playter. “This partnership combines our strong R&D foundations to tackle complex challenges and create robots that meet real-world needs.”

Gill Pratt, Toyota’s chief scientist and CEO of TRI, noted, “Recent advances in AI and machine learning have great potential to enhance physical intelligence. Merging TRI’s cutting-edge AI with Boston Dynamics’ hardware is transformative for both organizations as we aim to boost human capabilities and improve quality of life.”

The partnership will focus on rapidly developing whole-body behavior models for Atlas and other humanoid platforms TRI may work with. It will be interesting to see the new telepresence training systems, especially given Atlas’s complexity compared to TRI’s earlier bimanual setups.

Uncertainty Surrounds Atlas, While Others Push for Mass-Produced Humanoids

Boston Dynamics’ plans for Atlas as a commercial product remain uncertain. Meanwhile, companies like Tesla and Figure are designing humanoids for mass production, aiming to deploy thousands for simple tasks and gather data to improve swarm-based learning.

Tesla’s approach, with millions of cars collecting data, highlights that AI success relies on efficient data collection. Elon Musk envisions a machine that could replace most physical jobs.

While Boston Dynamics has led in humanoid development, Atlas remains a research platform. The company focuses on practical robots like Spot and Stretch, signaling humanoids may need more time before being ready for wide use. This evolving field feels like watching the future unfold.

Read the original article on: New Atlas

Read more: Lenovo’s Six-Legged Robotic Dog Unmatched Agility