Published on: February 25, 2025 / Updated on: February 25, 2025 – Author: Konrad Wolfenstein
From stumbles to somersaults in robotics: AI upgrade redefines humanoid capabilities – Image: Xpert.Digital
The future of humanoids: Atlas becomes smarter through reinforcement learning
Strategic partnership: Boston Dynamics optimizes Atlas for real-world applications
In an announcement, Boston Dynamics, a pioneer in dynamic robotics, and the Robotics & AI Institute (RAI Institute), a research institution led by renowned robotics expert and former Boston Dynamics CEO Marc Raibert, revealed a strategic partnership. The stated goal of this collaboration, which officially launched in February 2025, is to significantly enhance the capabilities of the advanced humanoid robot Atlas through the use of reinforcement learning. This collaboration promises not only to make Atlas more flexible and agile, but also to qualify it for a wider range of real-world applications, thus paving the way for a new era of humanoid robotics.
Suitable for:
Key objectives of the forward-looking collaboration
The partnership between Boston Dynamics and the RAI Institute focuses on a number of ambitious goals aimed at transforming Atlas's fundamental capabilities, evolving it from an impressive research demonstrator into a versatile and practical tool. These efforts center on three main areas:
Bridging the Sim-to-Real Gap: The Path from Simulation to Reality
One of the biggest challenges in robotics, particularly in the field of reinforcement learning, is transferring skills learned in simulations to the real world. Simulations offer an ideal environment for training robots because they provide unlimited data, complete control over the environment, and the ability to simulate dangerous or costly scenarios without risk. Robots can perform countless iterations of movements and tasks in virtual worlds without the danger of damage or injury.
Reality, however, is far more complex and unpredictable. Physical robots operate in a world full of sensory noise, unforeseen disturbances, inaccuracies in the modeling, and the constant challenge of variability. What works in a perfectly controlled simulation can fail in chaotic reality. The “sim-to-real gap” describes precisely this discrepancy.
The partnership between Boston Dynamics and the RAI Institute aims to close this gap through innovative methods and algorithms. Researchers are working to develop robust and generalizable motion sequences that function reliably not only in simulation but also in the real world. This includes developing advanced simulation environments that more accurately reflect physical reality, as well as employing techniques such as domain randomization and adaptive simulation to make the models trained in simulations more resilient to the unpredictability of the real world. Success in this area is crucial for unlocking the full potential of reinforcement learning for robotics and deploying robots in real, unstructured environments.
Improving Locomotive Manipulation: The Art of Movement and Interaction
The ability to loco-manipulate—that is, to move and manipulate objects simultaneously—is a key capability for robots intended to operate in complex and dynamic environments. Imagine a humanoid robot moving through a warehouse to pick packages, or a robot clearing debris in a disaster zone while simultaneously searching for survivors. In all these scenarios, it is essential that the robot can not only move efficiently but also interact with its surroundings at the same time.
However, developing advanced loco-manipulation strategies is an enormous challenge. It requires close coordination between motion planning, path planning, grasping planning, and force control. The robot must be able to adapt its movements and manipulations in real time to the constantly changing conditions of its environment.
As part of the partnership, researchers will develop new and innovative strategies to elevate Atlas's loco-manipulation capabilities to a new level. This includes exploring algorithms for simultaneous motion and grasp planning, developing robust force control strategies for manipulating various objects, and integrating sensor information into the control loop to enable responsive and adaptive loco-manipulation. Improving loco-manipulation is a crucial step in making Atlas a truly versatile and useful tool for a wide range of applications.
Exploring whole-body contact strategies: The synergy of arms and legs
Humanoid robots like Atlas have the unique potential to move and interact in ways that closely resemble human movement. This ability to integrate the entire body, including arms, legs, and torso, into complex movements and tasks opens up entirely new possibilities for robotics. Whole-body contact strategies go beyond simple arm manipulation and utilize the synergy between arms and legs to enable high-performance movements and tasks.
Imagine a person carrying a heavy object. They use not only their arms, but also their legs, torso, and entire body to stabilize the weight, maintain balance, and transport the object efficiently. Similarly, humanoid robots should be able to use their entire body to accomplish complex tasks that require close coordination between arms and legs.
The researchers are focusing on developing advanced control algorithms and planning strategies for high-performance whole-body movements and tasks. This includes areas such as dynamic walking, jumping, climbing, lifting and carrying heavy objects, manipulation in confined spaces, and interaction with complex environments. Research into whole-body contact strategies is crucial for realizing the full potential of the humanoid form factor and developing robots that can move and interact in the world in natural and intuitive ways.
The significance of this groundbreaking collaboration
The partnership between Boston Dynamics and the RAI Institute is of immense importance to the robotics and AI research community for several reasons. First, it unites two leading organizations in the field of robotics, each with unique strengths and expertise. Boston Dynamics is known worldwide for its impressive and dynamic robot platforms such as Atlas, Spot, Handle, and Stretch. The RAI Institute, under the leadership of Marc Raibert, brings decades of experience in developing cutting-edge technologies for intelligent machines and in applying reinforcement learning to complex robotics problems.
Marc Raibert, the founder of the RAI Institute, is an icon in robotics. As the former CEO of Boston Dynamics, he significantly shaped the company's development and created some of the world's most impressive robots. His vision of robots that can move in the real world with the same skill and versatility as humans and animals has profoundly influenced robotics research. With the founding of the RAI Institute, Raibert continues his mission to push the boundaries of what is possible in robotics and AI.
The collaboration builds on a solid foundation of previous joint projects, including the “Reinforcement Learning Researcher Kit” for the quadrupedal robot Spot. This kit enables researchers worldwide to develop and test reinforcement learning algorithms on the Spot platform. The successful development and implementation of this kit has demonstrated that both organizations are capable of working together effectively and developing innovative solutions in the field of reinforcement learning for robotics.
By applying reinforcement learning to Atlas, one of the world's most advanced and capable humanoid robots, the partners expect significant advances in the development of humanoid capabilities. Reinforcement learning offers the potential to train robots to handle complex tasks that would be difficult to achieve with traditional programming approaches. It enables robots to learn, adapt, and continuously improve their abilities through interaction with their environment.
Boston Dynamics and the RAI Institute have committed to publishing regular updates and demonstrations of their work with Atlas to make advances in humanoid robotics accessible to a wider public. This transparency is crucial for building trust in robotics and AI research and fostering public acceptance of these technologies. The planned publications will not only inform the scientific community but also inspire the public with the fascinating opportunities and challenges of humanoid robotics.
Joint research and development in detail
The collaboration between Boston Dynamics and the RAI Institute is divided into several core areas of research and development that are closely linked and complement each other:
Development of a shared reinforcement learning training pipeline for Atlas
At the heart of the partnership is the development of a state-of-the-art reinforcement learning training pipeline, specifically tailored to Atlas's needs and capabilities. This pipeline will form the basis for training dynamic and generalizable behaviors for mobile manipulation. It encompasses all steps of the reinforcement learning process, from defining reward functions and selecting suitable algorithms, through developing simulation environments and data acquisition, to validating and transferring the learned behaviors to the real robot.
The training pipeline will be modular to ensure flexibility and adaptability to different tasks and environments. It will integrate advanced reinforcement learning techniques, such as deep reinforcement learning, model-based reinforcement learning, and multi-agent reinforcement learning, to maximize training efficiency and robustness. A particular focus will be on developing reward functions that will enable Atlas to learn complex tasks without requiring every step to be explicitly defined. These reward functions will guide the robot to develop efficient, natural, and human-like movements and interactions.
Sim-to-Real Transfer: The bridge between the virtual and real worlds
As previously mentioned, the simulation-to-real-world transfer is one of the biggest challenges in reinforcement learning for robotics. The teams will work intensively to bridge the gap between simulations and the real world and to ensure that the behaviors trained in simulations can be successfully and reliably transferred to the physical hardware.
This requires a multi-layered approach that includes both improving simulation environments and developing robust transfer methods. The simulation environments are continuously improved to more accurately reflect physical reality, including the modeling of friction, contact, inertia, and other physical effects. Simultaneously, techniques such as domain randomization, system identification, and adaptive control are employed to make the models trained in simulations more resilient to the uncertainties of the real world. The goal is to create a seamless transition from simulation to reality, enabling Atlas to apply the skills learned in the virtual world to real-world environments without significant performance degradation.
Focus on key skills for the future of humanoid robotics
The partnership focuses on developing and improving key capabilities that are essential for the practical use of humanoid robots in real-world environments:
Improved locomotive manipulation: Handle objects while moving
Atlas should be able to manipulate objects and devices such as doors, switches, levers, tools, and other items while moving around. This capability is crucial for a wide range of applications, from industrial automation and logistics to search and rescue operations. Imagine Atlas navigating rugged terrain while simultaneously clearing debris or operating tools to repair a damaged structure.
Improved loco-manipulation requires the development of algorithms that coordinate motion planning, grasping planning, and force control in real time. Atlas must be able to adapt its movements and manipulations to the shape, size, weight, and texture of the objects it manipulates. Furthermore, it must be able to handle uncertainties in perception and the environment, dynamically adjusting its plans and movements. Developing these capabilities will make Atlas a far more versatile and useful tool for a wide range of applications.
Whole-body contact strategies: Complex movements and heavy loads
The researchers are focusing on developing sophisticated whole-body movements that go beyond simple walking and grasping. These include dynamic running, jumping, climbing, lifting and carrying heavy objects, and manipulation in confined spaces. These abilities require close coordination between arms, legs, and torso, utilizing whole-body synergy to accomplish complex tasks.
Dynamic walking and jumping enable Atlas to move quickly and efficiently across uneven terrain and over obstacles. Climbing extends its reach and allows access to hard-to-reach areas. Lifting and carrying heavy objects makes it a valuable tool in logistics and construction. Manipulation in confined spaces allows it to be used in environments that are difficult or dangerous for humans to access. The development of whole-body contact strategies is a crucial step toward realizing the full potential of the humanoid form factor and making Atlas a truly agile and capable robot.
Practical implementation and continuous progress monitoring
The partnership between Boston Dynamics and the RAI Institute places great emphasis on a transparent and practice-oriented implementation of their research and development work:
Regular progress reports and demonstrations
Boston Dynamics and the RAI Institute have committed to periodically publishing progress reports documenting the latest developments and achievements of their collaboration. These reports will include not only written descriptions of progress but also illustrative demonstrations using Atlas, showcasing the newly acquired skills in action. These demonstrations will be released as videos and presentations and made available to the scientific community and the general public.
The regular updates and demonstrations serve several purposes. They allow the scientific community to track advances in humanoid robotics and inspire one another. They promote transparency and trust in robotics research and help increase public acceptance of these technologies. Furthermore, they provide Boston Dynamics and the RAI Institute with the opportunity to receive feedback from the community and adjust their research direction accordingly.
Location of cooperation: Massachusetts, USA
All research and development work within the partnership takes place in Massachusetts, where both organizations are headquartered. This geographical proximity fosters close collaboration and direct exchange between the research teams. The Boston Dynamics and RAI Institute teams work in shared laboratories and utilize the resources and infrastructure of both organizations. This close integration of teams and resources is a crucial factor in the partnership's success, enabling the exploitation of synergies and the efficient advancement of research and development.
Atlas's Expected New Capabilities: A Look into the Future of Humanoid Robotics
Through the partnership between Boston Dynamics and the RAI Institute, the Atlas robot is expected to gain a range of groundbreaking new capabilities that will make it an even more versatile and useful tool:
Improved mobility and manipulation: Agility and precision in motion
Dynamic locomotion
Atlas will be enabled to move even more stably and smoothly across uneven terrain, in complex environments, and even in dynamic scenarios. This includes walking, jumping, climbing, and the ability to adapt to different surfaces and conditions in real time. Dynamic locomotion is made possible by advanced control algorithms and sensor data fusion, allowing Atlas to maintain its balance, overcome obstacles, and adapt its movements to the specific situation.
Full-body manipulation
The robot will implement advanced whole-body contact strategies to precisely and efficiently lift, carry, move, and manipulate heavy objects. This requires highly developed coordination of arms, legs, and torso to stabilize the weight, maintain balance, and handle the objects safely. Whole-body manipulation will allow Atlas to perform tasks previously reserved for humans, such as moving heavy loads in warehouses, on construction sites, or in disaster zones.
Enhanced environmental interaction: Intelligent interaction with the world
Object manipulation
Atlas will learn to manipulate a variety of objects and devices in his environment, including doors, switches, levers, valves, tools, containers, and much more. This ability will allow him to operate in human environments and perform tasks that require interaction with existing infrastructure. Object manipulation requires advanced perception skills to detect, locate, and identify objects, as well as sophisticated grasping and manipulation strategies to handle them safely and efficiently.
Adaptability to materials and structures
The robot will be able to automatically and intelligently adapt its force, speed, and movements to different materials and structures without damaging or destroying them. This is crucial for safe and reliable interaction in the real world, where robots will encounter a wide variety of surfaces, materials, and objects. This adaptability is achieved through the use of force and torque sensors, tactile sensors, and advanced control algorithms, enabling Atlas to monitor and adjust its interactions in real time.
Learning ability and generalization: The foundation for future innovations
More efficient learning through reinforcement learning:
By employing advanced reinforcement learning techniques, Atlas will be able to learn new skills significantly faster and more efficiently than before. This includes developing algorithms that accelerate learning and process data
Suitable for:
Your global marketing and business development partner
☑️ Our business language is English or German
☑️ NEW: Correspondence in your national language!
Konrad Wolfenstein
I would be happy to serve you and my team as a personal advisor.
You can contact me by filling out the contact form or simply call me on +49 7348 4088 965 (Munich) . My email address is: wolfenstein ∂ xpert.digital
I'm looking forward to our joint project.