A new machine learning framework called ZEST has enabled Boston Dynamics' humanoid robot Atlas to master complex full-body movements, including backflips and breakdancing, without requiring separate training for each individual motion. Developed jointly by the Robotics and Artificial Intelligence Lab (RAI) and Boston Dynamics, ZEST uses reinforcement learning, a technique where robots learn through trial and error, to train on diverse motion data sources simultaneously.

How Does ZEST Train Robots to Move Like Humans?

The breakthrough lies in ZEST's ability to combine multiple types of motion data in a single training process. Rather than teaching a robot one movement at a time, the framework ingests motion capture recordings of real human movements, video-based motion data, and keyframes from animation sequences. This diverse input allows robots to learn more generalizable movement patterns that can transfer across different actions without needing detailed instructions for each specific motion.

During training, ZEST automatically estimates the difficulty level of each movement and focuses more intensely on the harder motions to reduce failures. This adaptive approach helps robots learn efficiently by prioritizing the movements that require the most refinement. The research team published their findings in Science Robotics on August 12.

What Movements Can These Robots Actually Perform?

The results demonstrate a striking range of capabilities across different robot body types. Boston Dynamics' humanoid Atlas successfully executed movements including crawling on all fours, cartwheels, jumping onto boxes, ballet motions, handstands, and kicking a soccer ball. The robot also performed complex sequences that require simultaneous coordination of multiple body parts, such as crawling and breakdance routines.

Boston Dynamics' quadrupedal robot Spot, which has a fundamentally different body structure than Atlas, was additionally trained on animated motions and performed backward somersaults, maintained balance in a handstand posture, and executed sideways rolling movements. The research team also applied ZEST to Unitree's small humanoid robot G1, showing that the framework works across different robot designs.

How ZEST Changes Robot Development

• Unified Training Approach: ZEST eliminates the need to develop separate control techniques for each specific motion, reducing the engineering burden and accelerating robot development timelines.
• Generalization Across Body Types: The framework successfully transfers learning across humanoid and quadrupedal robots with different physical structures, suggesting broad applicability to future robot designs.
• Diverse Data Integration: By combining motion capture, video, and animation data in one training process, ZEST creates more robust movement models than single-source training methods.

The research team explained the significance of their work in a statement about the framework's potential impact. They noted that ZEST represents a major step forward in robot locomotion and could significantly reduce the engineering burden of developing separate control techniques for each specific motion.

"We expect ZEST to reduce the burden of having to develop separate control techniques for each specific motion and to help robots acquire a broader range of movements. With further technical improvements, it will contribute to closing the gap in movement capabilities between humanoid robots and real humans," the research team stated.

Robotics and Artificial Intelligence Lab (RAI) and Boston Dynamics research team

The significance of this work extends beyond impressive demonstrations. Humanoid robots are being developed specifically to perform various tasks in spaces designed for people, which means they need to move naturally and harmoniously like humans. Coordinating multiple body parts simultaneously to achieve fluid, human-like motion has been one of the most challenging problems in robotics. ZEST addresses this fundamental challenge by enabling robots to learn coordinated full-body movements from diverse real-world and synthetic motion data.

The framework's ability to generalize learned behaviors to a variety of actions without providing detailed information or separate reference motions for each movement represents a meaningful departure from traditional robot training methods. Rather than hand-coding control systems for each movement or training a robot repeatedly on each specific task, ZEST allows robots to develop a more flexible understanding of how to move their bodies across different scenarios. This approach mirrors how humans learn movement, by observing and practicing diverse motions rather than mastering one skill at a time.

As Boston Dynamics and other robotics companies continue to develop robots for real-world deployment, the ability to quickly teach robots new movements becomes increasingly important. ZEST's success with Atlas, Spot, and G1 suggests that future robot development could move faster and produce more capable machines. The framework's applicability to robots with different body structures also hints at a future where a single training approach could work across an entire ecosystem of different robot designs, from humanoids to quadrupeds to other specialized forms.