Our end-game is a universal robotics platform that is adaptable to a wide range of industrial jobs on- and offworld.


In our research and development of artificial intelligence solutions for real robotic systems we adhere to three core principles that define our focus on machine intelligence and embodied AI.


  • Let AI out of simulators into the physical world:
    The ultimate goal of AI and reinforcement learning is to train real agents to operate in the real world. To make progress in this direction we meet the challenge of applying this technology to real hardware.

  • Reality is more complex and rich than simulation:
    Simulators play an important role in the development of artificially intelligent robots, but at some point fidelity of even the best simulations becomes insufficient to prepare our robots for the real world. We can only simulate the known unknowns, which leaves unknown unknowns unexplored. We work with real systems and real environments to solve this issue.


  • Imprinting human intelligence into robotic units:
    Learning industrial operations is challenging not only for robots but also for human workers. Humans do not learn the job by trial and error, so why the robots should? Our approach and methodology allows expert human teachers to guide the robots and assist them during the learning process.

  • We replace hard-coded procedures with learned behaviors:
    Normally, it takes a team of engineers to describe the desired industrial operation in code, and this expensive effort produces a single-purpose procedure that is inflexible to changes in the environment. Learning behaviors from expert teachers saves time and engineering cost and can be modified at any time to meet new requirements and adapt to the whims of the environment.



  • Replace complex and outdated heuristics:
    We decompose robotic operation into subroutines and identify the ones that would benefit most from artificial intelligence and imprinted human intuition. We focus on the outdated and inflexible subroutines first, making sure our artificial intelligence solutions have an immediate and noticeable impact.


  • Build on top of well-optimized existing solutions:
    Over the years classical robotics has discovered amazing solutions to some of the problems. We integrate modern AI methods with well-proven classical algorithms, benefiting from the both worlds.


  • Widen the scope of what a robotic workforce is capable of:
    By expanding the capabilities of our industrial robotic workforce we expand the range of tasks that robots can tackle and open up new applications in mining, construction and other industries.


  • Bring robotic autonomy to a new level:
    Solutions that are based on artificial intelligence are by design more adaptable and robust to changes in the environment, uncertain and unknown situations and unforeseen challenges. A higher level of autonomy and reduced human involvement opens up the possibility of deploying robotic workforce in extreme and remote environments on Earth and in Space.



Visit our OffWorld Gym website:


  • A sandbox to raise robotic intelligence:
    Over the last years we have created a lunar analog environment to design and develop artificial intelligence and reinforcement learning algorithms in a controlled environment. We needed a testing area to evaluate the cutting edge research, identify which of the new methods proposed by the academia have applications in the real world, and teach our robots to make use of these methods. We have built such a testing area and called it OffWorld Gym.

  • The world-first open-access physical environment for reinforcement learning research:
    The ultimate challenge of reinforcement learning research is to train real agents to operate in the real environment, but until now there has not been a common real-world RL benchmark. The rest of the community who is working on the same problem has  exactly the same needs are we did. We decided to expand OffWorld Gym to include several physical training areas and open it to the public.

  • Direct the scientific community towards real-world problems in robot learning:
    The challenge that the community sets as a benchmark is usually the challenge that the community eventually solves. OffWorld Gym is a collection of real-world environments for reinforcement learning in robotics with free public remote access that serves as such benchmark. Close integration into the existing ecosystem allows scientists to start using OffWorld Gym without any prior experience in robotics. It removes the burden of managing a physical robotics system, abstracts it under a familiar API, and allows the researcher to focus on the algorithms. By solving real-world robotics challenges in our benchmark environments, we, together with the scientific community, generate solutions to those challenges. Once solved, they will have immediate applications in industrial robotics.




Himanshu Sahni, Toby Buckley, Pieter Abbeel, Ilya Kuzovkin
Addressing Sample Complexity in Visual Tasks Using HER and Hallucinatory GANs
Neural Information Processing Systems (NeurIPS), 2019

Ashish Kumar, Toby Buckley, Qiaozhi Wang, Alicia Kavelaars, Ilya Kuzovkin
OffWorld Gym: open-access physical robotics environment for real-world reinforcement learning benchmark and research
arXiv preprint, 2019

Deep Reinforcement Learning for Real-World Robotics
Talk at Artificial Intelligence in Robotics - Sydney, Australia - 2020

Machine Learning for Robot Autonomy
Booth presentation at International Astronautical Congress - Washington, DC, USA - 2019

Addressing Sample Complexity in Visual Tasks Using HER and Hallucinatory GANs
Poster at ICML workshop on Reinforcement Learning for Real Life - Long Beach, CA, USA - 2019

AI-Powered Industrial Robotic Workforce
Talk at Off Earth Mining Forum - Sydney, Australia - 2019

DRL for Robots in Extreme Environments
Talk at Deep Reinforcement Learning Summit - San Francisco, CA, USA - 2019