Boston Dynamics’ Robots Get Brains: Human Obstacles No Match

Boston Dynamics Goes End-to-End AI: Atlas Gets a Brain!

Boston Dynamics is leveling up its game, folks. This isn’t just about impressive robot gymnastics anymore. The latest upgrade to Atlas, their flagship humanoid, allows it to not only understand natural language commands, but to autonomously plan actions and handle unexpected disruptions.

In a demo showcasing this new capability, an operator deliberately closes a box’s lid. Not a problem for Atlas – it identifies the obstruction and opens the box without missing a beat.

Move the box? Atlas accurately recalibrates, factoring in the changed location.

And even if a crucial component is misplaced, Atlas can identify the missing piece and incorporate it into the task.

This upgrade, a collaborative effort between Boston Dynamics and the Toyota Research Institute, introduces Atlas MTS, powered by a Large Behavior Model (LBM). The partnership underscores the growing cross-industry collaboration needed to push robotics to the next level. The use of LBMs is key to giving these robots more intuitive interaction with their environment and tasks.

The robotics community is buzzing with anticipation. The official YouTube video showcasing the new capabilities has already garnered over 100,000 views and over 10,000 likes, highlighting the intense interest in the future of humanoid robotics.

One popular sentiment circulating online focuses on Atlas’s improved knee articulation, with commenters expressing relief that the design changes should lead to fewer back injuries for the robot. Concerns about robot ergonomics? Now *that’s* progress.

The company report provides detailed insights into the implementation of these new capabilities.

Wiring Up Atlas’s Brain

The report details how the end-to-end language-conditioned policy maximizes the robot’s inherent strengths, enabling walking, precise foot placement, squatting, and center-of-gravity adjustments while avoiding self-collisions. A critical component for any bipedal robot aiming for prime time!

Building this policy involves a four-step process:

Collecting embodied behavioral data;

Processing, labeling, and organizing the data;

Training the neural network;

Evaluating the policy with test tasks.

Of particular note is the Large Behavior Model itself, leveraging a 450 million-parameter diffusion Transformer model combined with a flow-matching objective. This architecture ingests data streams—30Hz images, proprioceptive data, and natural language instructions—and outputs motion directives to control Atlas. This is where the “brain” really kicks in.

Think of the Transformer as the “global eye,” managing the relationships across the robot’s architecture and motion details. Diffusion refines the movements, converting high-level commands into precise actions. And flow-matching loss anchors the movements in reality, ensuring actions are both realistic and adaptable.

Crucially, Boston Dynamics is coupling Atlas’s model predictive controller with a VR interface, addressing a wide array of tasks from fine motor skills to full-body movements and locomotion.

This allows remote operators to fully leverage the robot’s performance, synchronizing their own movements and perceptions with the robot’s state.

But the real game-changer? Atlas’s ability to autonomously handle unforeseen circumstances.

When issues arise – a dropped component, an unclosed box – Atlas can now intelligently react and adjust its behavior. This level of responsiveness opens doors to deployments in far more dynamic and unpredictable work environments, from construction sites to warehouses.

It’s essentially giving Atlas a brain on board!

Furthermore, the company stated that Atlas can learn any action demonstrated by a human, expanding their capabilities to include tying ropes, folding chairs, and flipping tires. All potentially useful in real-world scenarios.

This learning ability is impressive, suggesting a far easier time deploying and teaching Atlas new tasks!

Goodbye Hydraulics, Hello Electric: The Future of AI-Ready Robots

Speaking of Boston Dynamics, it’s important to remember the previous, hydraulically powered version of Atlas.

That version was retired in April 2024, with the all-electric version unveiled within 24 hours. A swift transition signaling a strategic shift.

Hydraulic systems, while powerful, are costly, have slower response times, and are challenging to integrate with AI systems. The shift is a testament to the need for robots to evolve in tandem with software advancements.

Electric drives offer greater precision, lower energy consumption, and a more natural fit within AI learning frameworks.

Since Boston Dynamics shifted to electric drive, they’ve been steadily rolling out new features.

Last August, Atlas showed off its push-up skills at RSS (Robotics: Science and Systems), a top robotics conference (even without fingers at the time, they could still perform a solid set of fist-pushups).

The form was surprisingly good.

Two months later, Atlas was ready for factory work. Give it location coordinates, and it could autonomously transport and sort objects.

And this past May, Boston Dynamics further upgraded Atlas by equipping it with 3D space perception and real-time object tracking, enabling it to perform even more intricate industrial tasks.

This latest Large Behavior Model solidifies their commitment to a radically new technological path. This suggests a significant upgrade to Atlas’s core software architecture.

Shifting gears from Boston Dynamics’ hydraulic-to-electric transition, let’s turn our attention to Unitree, which has consistently embraced electric drives.

From their quadrupedal Go series to humanoid robots like H1, G1 and R1, Unitree has adhered to a “small-and-beautiful” electric-drive philosophy, achieving rapid iterations and steadily increasing its global presence.

Now, they have even released a performance featuring 180 “ballet dancers.” How’s that for putting torque control to use?

Looking ahead, the true potential of electric-drive robots will emerge as electric drive technology and AI algorithms continue to mature and converge, heralding a new era in automation.