Improving Operability & Manoeuvrability of Canadarm2 & Dextre through Artificial Intelligence
Abstract — The Mobile Servicing System (MSS) of the International Space Station (ISS) is currently controlled by the Robot Operators (ROBOs), a NASA & CSA joint team of operators operating simultaneously from the ground operation bases in St-Hubert and Houston. Whilst this frees up astronauts from having to control the MSS themselves, the cost on the ground is great. ROBO missions are long and arduous, oftentimes taking multiple days. They are frequently interrupted by loss of connection, require extensive calculations by the operators, and are subject to human errors. The use of Machine Learning (ML) and computer vision could drastically improve the operability and manoeuvrability of the MSS by supporting the operators in their duties.
1 INTRODUCTION
Artificial Intelligence (AI) is gaining more and more traction in multiple fields, increasing automation and solving complex problems that were previously only possible for humans. AI and the use of big data can lead to tangible advancements in the operation of space assets, improving coordination and efficiency in task planning, as well as allowing space assets in deep-space to react autonomously in response to their environment.
The Mobile Servicing Systems (MSS), comprised of the Canadarm2 and Dextre, are sensor-enabled robotic arms that are used for servicing tasks aboard the International Space Station (ISS). Both are remotely controlled by the Robot Operators (ROBOs), a NASA & CSA joint team of operators operating simultaneously from the ground operation bases in St-Hubert and Houston.
Over the past 17 years, the Canadarm2 has serviced the ISS and other space assets with resounding success. However, this resounding success is not without its drawbacks: Due to the human, ground-controlled nature of the system, the MSS is prone to human errors and require extensive, arduous operations from the personnel in the ROBO team. With the improved accessibility and popularity of space platforms, this problem will only become more significant in the years to come. Indeed, the ISS is faced with an increase in servicing requests from public and private entities [https://www.nasa.gov/johnson/HWHAP/robotic-arms-in-space], and their successful operations only galvanize the demand more.
In order to satisfy to the increasing demand, the CSA and NASA will have to look for more efficient and autonomous solutions. In this report, we will demonstrate that the use of Artificial Intelligence in the control process of the MSS, the ROBO could drastically improve its efficiency, operability and manoeuvrability by supporting the operators in their duties.
We will first begin by laying down the necessary theoretical notions to understand the scope of this document. We will then move on to evaluate the current alternatives and landscape, both in the space industry and in relevant similar fields. With this context in mind, we will propose our solution, detailing its roadmap to completion, important steps, effectiveness to solve the problem, innovation ability, and technology readiness level. Lastly, we will explain how we intend to mitigate the potential risks within Phase 1 and assess the technical feasibility of the solution.
2 THEORY
2.1 Artificial Intelligence Theory
Artificial Intelligence (AI) — AI can be defined as “the study of [intelligent agents, or] any device that perceives its environment and takes actions that maximize its chance of successfully achieving its goals” [https://en.wikipedia.org/wiki/Artificial_intelligence].
Machine Learning (ML) — Arthur Samuel, pioneer in artificial intelligence, coined the term “machine learning” in 1959. He defined machine learning as the “field of study that gives computers the ability to learn without being explicitly programmed” [Samuel, A. L. (1959), “Some Studies in Machine Learning Using the Game of Checkers” in IBM Journal of Research and Development (Volume:3, Issue: 3), p. 210]. Nowadays, ML is also commonly used to define a set of techniques used in AI to solve intelligent-agent types of problems.
Deep Learning (DL) — Deep learning is a subset of ML which relies on Neural Networks (NN), and exhibit the ability to learn patterns from structured and unstructured data alike. The vast majority of deep neural network are trained using labeled data .[https://www.investopedia.com/terms/d/deep-learning.asp].
Neural Networks (NN) — Neural Networks are mathematical structures composed of dozens of layers of matrix operations. These are especially efficient at extracting complex patterns and information out of large, labelled datasets. For instance, neural networks can classify images in thousands of categories, translate languages automatically, detect words from sound waves, etc.
Reinforcement Learning (RL) — RL is a technique used to train deep neural networks through a scoring system to train an intelligent agent. Points are awarded to the agent for encouraged behaviours, whereas penalties are given for discouraged behaviours. These points are then fed to the neural network as feedback on its internal state. The neural network uses this feedback to tune its million of internal parameters to maximize the probability of getting the right output. Think about a dog being rewarded with a biscuit for intended behaviours and reprimands for bad behaviours. Over time, the dog will learn to expect certain rewards and punishments and will act in order to maximize the rewards.
Computer Vision — Computer vision is a field specialized in methods used for “acquiring, processing, analyzing and understanding digital images, and […] transformation of visual images […] into [high-dimensional data about] the world [used to feed learning algorithms and extract intelligent decisions]” [https://en.wikipedia.org/wiki/Computer_vision].
Convolutional Neural Networks (CNN) — CNN are neural networks specifically dedicated to Computer Vision tasks through the use of Deep Learning.
2.2 Startup Methodology Theory
Minimum Viable Product (MVP) — Eric Ries, author of the Lean Startup, often cited as a foundational text of the startup industry, defines an MVP as “[the] version of a new product which allows a team to collect the maximum amount of validated learning about customers with the least effort” [https://hackernoon.com/the-ultimate-guide-to-minimum-viable-product-59218ce738f8]. In other words, the MVP is “all about solving the problem by the most basic solution” [https://hackernoon.com/the-ultimate-guide-to-minimum-viable-product-59218ce738f8]. This mindset has great benefits, such as:
- Faster testing
- Faster learning
- Minimizing R&D costs
3 CURRENT LANDSCAPE
3.1 Current Solution & Issues
As expressed previously, the Mobile Servicing Systems (MSS) on the ISS is primarily jointly controlled by the ground control centers, located in St-Hubert and Houston respectively. The Robot Operators are tasked with extremely complex, often very precise tasks, and are frequently interrupted by loss of connection [https://www.nasa.gov/johnson/HWHAP/robotic-arms-in-space]. This mix of constraints can lead to human errors and very long and arduous mission schedules.
3.2 Alternatives
Currently there does not exist any alternatives to onboard-control and remote, ground-control of the MSS. All operations have to be handled by highly trained teams of operators, which must go through extensive simulation training [https://www.youtube.com/watch?v=ck76yL_s30s] and have a strong background in the understanding of physics, mathematics, mechanical engineering, and many more fields.
It is very difficult in this context to find alternatives; the Canadarm2 and Dextre are unique in their field. There is no other robotic (controlled or automated) components currently used in space.
3.3 Similar Efforts in Related Fields
The situation is different when we take a look at what is done in other industries. In this section, we will review similar efforts done in the fields of industrial factory & assembly line robots, and self-driving cars.
3.3.1 Factory & Assembly Line Robots
In 2012, Amazon’s warehouses were primarily coordinated and operated manually by thousands of human employees. In that same year, the giant acquired Kiva for US $775M [https://qz.com/709541/amazon-is-just-beginning-to-use-robots-in-its-warehouses-and-theyre-already-making-a-huge-difference/]. Starting Q3 2014, they started deploying their own AI-powered warehouse bots that coordinate between each other. Within a year, the “click to ship” cycle decreased from between 60 and 75 minutes to 15 minutes [https://qz.com/709541/amazon-is-just-beginning-to-use-robots-in-its-warehouses-and-theyre-already-making-a-huge-difference/]. And this is not a small feat, considering that manipulating objects is still one of the most complex challenges in robotics today [https://www.wired.com/story/grasping-robots-compete-to-rule-amazons-warehouses/].
Whilst the use of machine learning in manufacturing is a quite recent field, analysts forecast that AI-powered robots will increase production up to 20%, even in cases where tasks are only partially automatable [https://www.mckinsey.de/files/170419_mckinsey_ki_final_m.pdf]. Moreover, it is expected that the semiconductor industry could reduce their scrap rate by up to 30% by using AI-enabled robots [https://www.mckinsey.de/files/170419_mckinsey_ki_final_m.pdf].
These are significant signs that machine-learning powered robots can lead to great improvements in efficiency and accuracy. This applies to both complex, coordinated tasks like Amazon’s, and simpler tasks like repetitive manufacturing chain tasks.
3.3.2 Self Driving Cars
Self-driving cars are a brand new field of artificial intelligence, with less than a decade of R&D behind it. Over the past five years, companies such as Tesla Motors, Waymo, Google, General Motors and Uber have shown an interest in the technology, pushing the field forward.
While it is still unclear whether self-driving cars are currently safer than human-driven cars, what’s clear is that they are improving very fast. Self-driving cars have to deal with very complex environments with a high number of independently moving agents. They must use AI and Computer Vision to make sense of the environment around them, and make decisions based on their surroundings. They oftentimes have to deal with great deals of uncertainties from the other human beings in their surroundings.
In that sense, self-driving cars could be said to evolve in even more complex environments that that of the MSS in space, due to the uncertainty arising from human agents. By comparison, outer-space is a much more predictable environment for an AI to evolve in. Indeed, with a sufficient amount of sensors, it can calculate and predict with accuracy the future states of the environment and adapt consequently.
If the technology underlying self-driving cars is sufficiently advanced to be allowed to roam freely on our roads without grave concerns for the lives of other drivers, then the technology is most likely sufficiently mature to allow R&D in the context of the MSS.
3.4 Conclusion of Assessment of Alternatives
In conclusion AI-enabled robotic is still in its nascent stage. Great amount of progress and results has been shown in related industries such as industrial robots and self-driving cars.
By implementing the proposed solution, we would not only drastically improve the current situation aboard the ISS, but also we would pioneer the entire field, and potentially bring ground-breaking innovations to the field of autonomous intelligent agents, both on Earth and in space.
4 PROPOSED SOLUTION
4.1 Proposed Solution: An Abstract
We propose to improve the operability, manoeuvrability and efficiency of the Mobile Servicing System (MSS) through the progressive introduction of artificial intelligence in the control process of the said system.
We expect to do so through the use of Deep Learning and Convolutional Neural Networks (CNN) to achieve highly accurate and efficient Computer Vision, coupled with Reinforcement Learning to teach the AI how to successfully complete its task.
We will train the artificial intelligence to do specific tasks using data from the previous missions as well as simulated training environments. We will train the AI to become as efficient as possible whilst optimizing for safety and positive outcomes.
These rewards and penalties will be given based on the ROBO team’s understanding of what actions are considered positive actions (safe maneuvers, valid actions, etc.) and negative actions (jerky movements, release of large momentum, overuse of force, invalid actions for the problem at hand).
4.2 Early Areas of Improvement
Automating the MSS is no small feat. The MSS is a highly complex system, evolving in a complex, unforgiving environment. It also undertakes tasks of high complexity.
In the Minimum Viable Product (MVP) mindset, we propose to start small, and grow larger as we validate and implement each module. Here are some early, self-contained problems that could prove to be good candidates to begin with:
- Improve the automatic docking sequence of the Robotic Arm Latching End Effectors (LEE — essentially the hands of the Canadarm2).
- Implement a process that would optimize the trajectory and speed of movement of the MSS around the ISS, including preemptive collision detection algorithms.
- Automatize routine servicing tasks like electronic box replacements
4.3 Roadmap to Automation
We understand the risks that can arise out of automation. That is why we propose a gradual adoption of the technology, at a pace that ensures constant safety and risk management.
As of September 2016, the US Department of Transportation adopted “The 5 levels of driving automation” adoption plan for self-driving cars [https://www.vox.com/2016/9/19/12966680/department-of-transportation-automated-vehicles]. We propose to adapt the same plan to the gradual automation of the MSS.
4.3.0 Level 0: No Automation
This is the current situation. There is no current automation of the MSS.
4.3.1 Level 1: ROBO Assistance
We propose to assist the ROBO team in their operations by providing cues on the optimal commands obtained by the AI. These cues are only there for informational purposes, and the operators may decide or not to implement them.
4.3.2 Level 2: Partial Automation
We propose to automate some smaller, less complex and very repetitive tasks, such as those presented in §4.2.
These automated actions may or may not be carried under the supervision of the ground control, based on their decision.
4.3.3 Level 3: Conditional Automation
At this level, we propose that the ML algorithm gains some automation features under certain specific conditions. In the context of self-driving cars, this would mean that the AI gains the ability to react when it foresees potential accident ahead of time. In the context of the MSS, that would mean to prevent foreseeable undesired consequences, such as sudden release of large momentum, collision with other parts of the ISS and MSS, etc.
These automated actions may or may not be carried under the supervision of the ground control, based on their decision.
4.3.4 Level 4: High Automation
At this level, we propose that the AI becomes in charge of routine operations such as resupply missions and servicing of electronic boxes. More complex missions, such as tailored servicing mission for third-party space assets (private and public) would be only partially automated. The most complex parts of these missions would be completely handled by the ground control.
Once again, the automated actions may or may not be carried under the supervision of the ground control, based on their decision.
4.3.5 Level 5: Full Automation
At this level, we propose that the AI becomes fully in charge of all missions carried out by the MSS. These include but are not limited to: routine missions, complex, tailored servicing missions. By this point, the ground operators would be mainly responsible for devising and communicating the necessary information for the AI to grasp the goal of the mission as well as the important actions to be undertaken.
At this level, the ground operators still retain the ability to gain full control of the MSS at any moment. However at that point, the likelihood of such a need to arise is low.
4.3.6 Scope of Mandate
We propose to achieve Level 3 of Automation of the MSS by the end of the given period for Phase 2 (2020).
Further automation will have to be evaluated with the CSA and NASA based on the space landscape and their objective by then.
4.4 Demonstration of Effectiveness
Artificial Intelligence can find solutions that trumps even the best minds. As designer and futurist Maurice Conti shows in his TED talk, Artificial Intelligences work through possibilities via an evolutionary process. They thus learn to get more efficient every time they make mistakes, just like nature would with evolution [https://www.ted.com/talks/maurice_conti_the_incredible_inventions_of_intuitive_ai]. As he further points out, AIs can come up with designs that are extremely alien to a human mind, yet orders of magnitude better [https://www.ted.com/talks/maurice_conti_the_incredible_inventions_of_intuitive_ai].
The online class Artificial Intelligence A-Z™ by SuperDatascience on Udemy provide a very simple example to understand how AI can take better decisions than humans [https://www.udemy.com/artificial-intelligence-az/]:
Let’s say you decide to put an AI into a labyrinth layered with traps. Its goal: avoid traps and reach the exit. Let’s also say that whenever the AI decides to walk forward, it has 80% of moving forward like intended, and 10% chance to move to either side.
In this situation, whenever the AI is confronted to a situation in which it is next to a trap, it will choose to bump itself in an opposing wall forever over risking to fall in the trap. As you can see in Figure 3, the green arrows denote the optimal choice made by the AI given the situation explained above.
As much as this is an odd plan of action, this situation is effectively an optimal one. Indeed it removes the risk of falling into the trap. A normal human would not think of this.
This very simple case shows how an AI can take unexpected, but better decisions that its human counterpart. What’s more, this ability exponentiates as the complexity of the environment increases, thus potentially making an ML-powered MSS much more efficient and safe than a remote, human-controlled one.
4.5 Innovation Ability of Solution
The use of Artificial Intelligence in space is a brand new field, with very little research done on the subject. As it is, NASA started to look into using AI in space only last year by founding its Frontier Development Lab (FDL), an accelerator dedicated to exploring the use of AI in space [http://www.frontierdevelopmentlab.org/].
While that alone is sufficient to make our proposal innovative, it does not stop there. The use of Computer Vision and Deep Learning to make sense of complex environments is also a field that has only bloomed in the last few years. The pioneers of this domain are military contractors such as DARPA, and self-driving car companies. AI in robotics, especially when it comes to learning to perform complex tasks with physical hardware such as robotic arms, is a very young field.
As you may already know, Montreal is the world leader in artificial intelligence, with the Montreal Institute for Learning Algorithms (MILA) at its center. We are lucky to have the support of Matthieu Courbariaux, member of MILA and PhD student in Deep Learning under the supervision of Yoshua Bengio. His help, alongside our partnership with District3’s Innovation Program will allow us to push the use of artificial intelligence beyond and potentially advance the science in the field.
This void in the market, coupled with the high expertise of our team is bound to help us achieve a forefront position in the market for automats in space.
4.6 Demonstration of Technology Readiness Level (TRL)
The underlying technology for this project has been battle tested, and can be inserted in the TRL 6–7. Indeed, Artificial Intelligence, Machine Learning, Deep Learning, Neural Networks and Computer Vision are all fields that have proven usability in real world contexts. Take for instance self-driving cars; they rely on most of the technologies listed above and are tested every day in the streets of major cities all around the United States.
This however does not mean that the proposed solution itself is also of a TRL 6–7. Indeed as explained in the previous section, the complete void of competitors regarding the use of AI in space robotics, and youth of related fields like the use of ML in manufacturing robotics clearly reduces the TRL of this project to the level 1 or 2.
Moreover, since no solution for the proposed problem has been developed yet, this proposition is simply in the stage of early ideation, further supporting that the TRL of this project is 1 or 2.
Finally, no further research and development than that contained within this document has been done prior to the application to the Artificial Intelligence and Big Data Analytics for Advanced Autonomous Space Systems challenge.
4.7 Risk Assessment
4.7.1 Lack of Interest from the Canadian Space Agency
While quite simple, this is the greatest risk any project faces. This is why we leverage the business and startup methodology expertise of District3 as well as Arthur Intelligence’s in order to validate the needs of the Canadian Space Agency early.
We intend to address and mitigate this risk early on in the Phase 1 by running a series of business model validation interviews with the CSA employees and upper staff to determine the direction that it wants to go. We will adjust our direction to match that of the CSA and better serve its needs.
In the event that the CSA does not have an interest in this proposed project nor any of the two other, we commit to pivot and adapt to build another project that suits better the CSA needs.
4.7.2 Safety & Maturity of Technology
Whilst Machine Learning is not a recent breakthrough in Computer Science, applying Deep Neural Network and Computer Vision to complex environments is a relatively new field. The technology is yet to be fully mature, and has not been fully polished.
In a complex and dangerous environment like that of the ISS, mistakes are costly, and can easily lead to the death of astronauts. This is why we have opted for the standardized model of the Automation Levels. This allow gradual adoption of the technology whilst eliminating risks by testing the algorithm under the express supervision of the ROBO team. Before moving to higher levels of the Automation Levels, the algorithm will have to pass the industry standard of Six Sigma, or 99.99966% reliability.
Whilst this risk cannot be mitigated during Phase 1, it is a risk that is continuously mitigated throughout the development, manufacturing and implementation process. Thus the mitigation will be implemented over the course of the Phase 2 and in iterations beyond the Phase 2.
4.7.3 Insufficient Training Datasets
Machine Learning, and especially more advanced techniques like Neural Networks require large bodies of data in order to train correctly an algorithm to perform complex tasks. This is why self-driving car companies send their vehicle throughout USA with human drivers; with this large collection of road experience, it is possible to create much more accurate models of the task to be performed.
Likewise, the ML algorithm to be employed on the MSS to increase automation will require a large database of experiences in order to become useful. We intend to rely on the datasets from previous missions to fulfill this need of data. Due to the large number of cameras and sensors on the MSS and ISS, we would expect the CSA and NASA to have a large amount of data readily available to train our models.
If however this is not the case, we intend to mitigate the effect of this issue by training our Neural Network primarily within simulations. We expect the CSA and NASA to already have scientifically accurate simulators for training of operators as well as for developing new servicing missions like with the Robotic Refueling Mission (RRM), carried by the NASA and CSA in 2013 [https://sspd.gsfc.nasa.gov/rrm_refueling_task-blog.html]. The RRM mission was the ISS’s first mission which goal was to refuel and service a private space satellite. The mission required extensive simulations and real-life training with training payloads similar to the satellite to-be-serviced. The real-life training took place both with high fidelity replicas on earth and in space [https://sspd.gsfc.nasa.gov/rrm_refueling_task-blog.html]. We expect that if the NASA and CSA were ready to train their operators in real-life situations and develop high-cost replicas for a single mission, then they would be ready to provide the same levels of efforts to train the MSS to become more autonomous.
It is important to bear in mind that the quality of the simulations will greatly impact how well the neural network will perform in real situations. The closer the simulations will be to reality, the better the algorithm will be.
We intend to address this risk during Phase 1 by establishing the extent to which we have access to trainable data, and to which the NASA and CSA are ready to invest into training the ML algorithm set to power the MSS.
As a measure of safety, the real-life training of the MSS would be undertaken under the supervision of the ROBO team, and all actions of the ML algorithm will have to be accepted by the ROBO team before being transmitted to the MSS.
4.7.4 Insufficient Input Sensors on MSS
Even given large datasets from previous missions and large budgets for real-life training of the algorithm, there is a possibility that the MSS does not have sufficient onboard sensors to gather data that can be used for training.
We do not believe this to be a high risk, given that the MSS is equipped with a large amount of onboard electronics, including force-end moment sensors as well as boresight, on-arm and on-ISS cameras.
With this in mind, the worst that could happen is that the cameras would have too low of a resolution for the Convolutional Neural Network (CNN) to pick up patterns clearly.
We intend to mitigate this risk by training our algorithms on the available datasets and verify early-on that key patterns can be recognized by the neural network. In the event that this cannot be done, we intend to enter in a discussion with the NASA and CSA to establish if they would be willing to upgrade their sensors (mainly cameras) to higher definitions.
4.7.5 Insufficient Computational Resources
Machine Learning and Deep Learning exist since the early 50’s and 70’s respectively, however both of these technologies have only reached mainstream usage in the last years. This is mainly due to a lack of processing power. Indeed it is only in recent years that computers have reached the necessary performance to perform the computations necessary for ML and DL.
It is no wonder then that the MSS were not designed with this technology in mind. As
Tim Braithwaite, the Canadian Space Agency’s Liaison Manager, formely member of the ROBO team states, “even a few years ago, [using AI on the MSS] wasn’t even conceivable” [https://www.nasa.gov/johnson/HWHAP/robotic-arms-in-space]
It follows then that we are faced with a complex problem: If computers back in 2001 (year of installation of the Canadarm2) were not sufficiently powerful for ML, then how can we retrofit them with such capabilities?
Matthieu thesis is specifically directed at making Deep Neural Networks more computationally efficient. His help in Phase 1 and 2 of the project will help us mitigate this risk with state of the art research.
However there is no guarantee that our best efforts succeed to fit the program on the onboard computers of the MSS. This is especially true given that embedded computers, like those found in satellites and robots, are oftentimes optimized to be as resource efficient — and therefore less performant — as possible.
In all cases, we suggest that throughout Automation Levels 1–3, most of the computations be rendered in the ground control center. While this solution does not improve on the loss of communication problem, it still delivers on all other aspects of the problem. Later phases will have the possibility to be carried out on new, more performant computers sent to the ISS during a resupplying mission.
4.8 Technical Feasibility Analysis
Even with proper mitigation of the risks analyzed above, it is important to analyze the technical feasibility of the proposed solution. More specifically it is important to delimit which part of the solution are the most complex and take longer to develop.
In today’s open source environment on the Internet, we are lucky to have access to countless resources in the following forms:
- open source code repositories
- open source databases
- previously attempted projects
- learning resources
- online tutorials
and most importantly, academic publications as well as arxiv tech reports.
And AI does not make exception — quite the opposite. By leveraging commonly used open-source libraries such as TensorFlow, Torch, pandas, scipy, etc., we can greatly improve our team productivity and deliver a product within much shorter deadlines. For instance, an author on Medium named Max Deutsch produced a proof-of-concept self-driving car with steering-wheel abilities in as little as 26 hours by leveraging different codebases available online [https://medium.com/the-mission/how-to-build-a-self-driving-car-in-one-month-d52df48f5b07].
Then, our biggest hurdle will not come from building a viable artificial intelligence program, but rather from acquiring sufficiently high quality data from the datasets as well as the simulations to make them usable by the AI.
We will also need to interface with proprietary systems within the CSA and NASA in order to obtain the best quality data possible. This may prove to be a major technical challenge; legacy and proprietary solutions are not always simple to interface with.
Most of the time will be spent on analyzing the datasets that we have access to in order to extract quality trainable data for our machine learning algorithm. In effect, most data scientists spend up to 80% of their time obtaining, cleaning and preparing their datasets for later modelling. The quality of the data we get from the CSA and NASA will directly affect the performance of our algorithms as well as the time necessary to build the solution.
We intend to mitigate this risk by partnering with the IT professionals and engineers behind the systems at the CSA and the NASA. By leveraging their expertise of the system, we will be able to better address the need and build a better solution.
5 CONCLUSION
In conclusion, through the use of Deep Learning, Convolutional Neural Networks (CNN), Computer Vision and Reinforcement Learning, we believe that we can drastically improve upon the existing situation when it comes to operating the MSS. We believe that an automated MSS could achieve highly accurate maneuvers and help the ROBO team optimize their mission planning as well as relieve them of recurring tasks.
As shown earlier, there is currently a void in the niche of space robotics and the use of AI in space, and we firmly believe that our solution would be first-of-its-kind and improve upon the current state-of-the-art, both in the space niche and the overall use of automated robots to carry on complex tasks.
It followed that we demonstrated that whilst our solution bases itself on more technologically ready technology, our solution itself fits well within the prescribed TRL 1–4.
In order to optimize for safety and risk management, we proposed to implement our solution via the industry standard of the five levels of automation. We have also shown that existing situations in related field show that ML and AI are great candidates for autonomous agents evolving in complex environments, even coming up with solutions better than what humans could come up with.
We demonstrated that we have carefully evaluated the risks that our solution faces, and that we have planned for these risks through the use of our resources and startup experience.
In the end, we believe that the use of Machine Learning (ML) and computer vision could drastically improve the operability and manoeuvrability of the MSS by supporting the operators in their duties. This is especially true in the current context of the industry.
You can find the full PDF version of this proposal here.
